2010

Society for Mathematical 43rd Annual Conference Portland, OR

Program & Information Welcome Presentation Guidelines

Welcome to MP2010, the 43rd annual meeting of Talks the Society for Mathematical Psychology, being Talks will be held in three parallel sessions, in held in the Embassy Suites Portland in Down- the Gevurtz, Fireside and Queen Marie room. town Portland. Please see the schedule at the end of this pro- This year’s conference features plenary ad- gram. Presentation time will be limited to a to- dresses from Mike Jordan, Larry Maloney, and tal of 20 minutes, which includes five minutes for William K. Estes Early Career Award winner discussion. Talks will be strictly timed, and syn- Thomas Griffiths. There are 4 invited symposia, chronized across sessions. 122 accepted talks and 22 poster presentations. We ask the last presenter in each session to keep time (since they are the person most mo- Registration tivated to keep to the schedule!), and arrange a time-keeper for their own talk. You can register from 6pm–9pm in the Colonel Lindbergh room (overlapping with the reception Posters from 7pm–9pm), or during the conference. You will receive a printed copy of this program, The poster session will be held Sunday from including the abstracts and schedule, as well as 5pm–6.30pm in the Colonel Lindbergh room. your name tag, which is stamped to indicate ban- The poster boards are 8’ wide by 4’ tall. quet purchases. Acknowledgments Conference Events Sponsors Reception We are grateful to the Air Force Office of The reception is Saturday evening, from 7pm– Scientific Research (AFOSR), and the Na- 9pm in the Colonel Lindbergh room, overlapping tional Science Foundation (NSF) for their with registration. There will be a cash bar. generous financial support for the conference.

Business Meeting We also thank the Cognitive Science Society for sponsoring the prize for the best student The business meeting will be held from 5pm poster on a cognitive modeling topic. Monday, in a location to be announced. People Conference Banquet Conference Chair: Robin Thomas The conference dinner will be Monday from Local Organizer: Misha Pavel, Sandy Baxter 7pm–10pm in the Queen Marie Ballroom. Tick- Financial Chair: Rich Golden ets for the banquet need to pre-purchased. Academic Program : Michael Lee Webmaster: Annemarie Zand Scholten Lunch Website 2010: James Pooley Registration: Jenny Shi, Daniela Amado Morning & afternoon coffee breaks are provided Business Consultant: Pernille Hemmer during the conference, but lunch is on your own.

Program Booklet Credits LATEX-code generation, Timo Kluck, Infty Advies lay-out, cover: (www.infty.nl) with MLC, inc.

1 Abstracts

INVITED PLENARY TALKS Berkeley . Bayesian inference is often viewed as an assumption-laden approach to statistical in- Sunday, 13:45 (Session B) ference, whereby strong assumptions about ex- Movement Planning under Risk, Decision isting knowledge are imposed in order to support Making under Risk. Larry Maloney NYU the acquisition of new knowledge from data. The . In executing any speeded movement, there is past two decades have seen the development of an uncertainty about the outcome due to motor increasingly vigorous field of Bayesian nonpara- variability. I’ll describe recent experiments in metrics, which simultaneously provides a more which subjects carried out speeded motor tasks expressive language for prior knowledge and al- that were equivalent to economic decisions. The lows for weaker specifications. Mathematically, outcome of each movement earned an explicit Bayesian nonparametrics amounts to using gen- monetary reward or penalty. In one game, for eral processes as prior distributions. example, subjects attempted to reach out and I discuss a class of stochastic processes known touch briefly presented reward disks while avoid- as “completely random measures” that I view as ing nearby, overlapping penalty disks. The task providing particularly useful building blocks for for the subject was to trade off the risk of miss- Bayesian nonparametrics. I will discuss appli- ing the reward disk against the risk of hitting the cations of models based on completely random penalty disk. The optimum tradeoff depended measures to several problem domains, including on the magnitudes of penalty and reward and natural language processing, protein structural the subject’s own motor error. For each sub- modeling, computational vision, and statistical ject, in each game, we could estimate the ideal genetics. movement strategy and maximum expected gain possible. We compared subjects’ movements and Tuesday, 13:45 (Session B) winnings to this ideal. In many of these exper- Using Probabilistic Models of Cogni- iments, subjects consistently chose movements tion to Identify Human inductive Biases. that were close to optimal. This outcome is sur- Thomas Griffiths UC Berkeley . People are prising: these motor tasks are, after all, formally remarkably good at acquiring knowl- equivalent to decision making under risk and sub- edge from limited data, as is required in learn- jects making decisions under risk typically do not ing causal relationships, categories, or aspects of maximize expected gain. I’ll describe very recent language. Successfully solving inductive prob- work in which we out to translate classical de- lems of this kind requires having good “induc- cision making experiments (concerning the inde- tive biases” – constraints that guide inductive pendence axiom of expected utility theory) into inference. Viewed abstractly, understanding hu- motor form and compare decision making under man requires identifying these inductive risk to movement planning under risk in the same biases and exploring their origins. I will argue subjects. that probabilistic models of cognition provide a framework that can facilitate this project, giv- Monday, 13:45 (Session B) ing a transparent characterization of the induc- Recent Developments in Nonparametric tive biases of ideal learners. I will discuss three Bayesian Inference. Mike Jordan UC ways in which these models can be used to shed

2 light on human inductive biases: comparing pre- meanings can be learned. All languages have dictions, assessing learnability, and designing ex- strong syntactic-semantic correlations. The se- perimental methods that magnify the effects of quential order in which words occur, the argu- inductive biases. ment structure and general syntactic relation- ships within sentences, all provide vital informa- tion about the possible of words. In this INVITED SYMPOSIA work, we will describe a set of Bayesian distribu- tional models that go beyond the “bag-of-words” New computational approaches to word paradigm. These models provide a framework learning. (Organizer: Amy Perfors) to address how semantic representations can be learned from coarse and fine-grained distribu- Tuesday, 9:00 (Session A) tional , including those provided by se- New computational approaches to word quential and syntactic information. They also learning. Luke Maurits, Amy Perfors, allow us to address the key issue of how polyse- Daniel Navarro, University of Adelaide . All mous words are represented and learned. In par- language learners must face the task of infer- ticular, we will address how distributional models ring the mappings between words and their ref- erents. Similarly, learning which word orders are that incorporate sequential and syntactic infor- mation explain how these continua of can permitted in one’s language is one of the first grammatical learning tasks those same learners be learned and disambiguated. must solve. We present modelling results sug- Tuesday, 9:40 (Session A) gesting that: 1) word order distributions across Beyond Boolean Logic: A learning the- languages may reflect an optimal response to the ory for complex compositional concepts. computational problem of conveying information Steven Piantadosi, Joshua Tenenbaum, over a noisy channel, as the Uniform Informa- Noah Goodman, MIT . We study the learn- tion Density (UID) hypothesis suggests; and 2) ing of novel words which express set-theoretical learning is improved when word order and word concepts often found in the functions words of reference are acquired jointly. natural language. We first present an experi- Tuesday, 9:20 (Session A) ment in which subjects inferred the meanings of Learning Word Meanings from Sequential a word which denoted an abstract concept that and Syntactic Statistics. Mark Andrews, could only be expressed in a rich representational NTU and UCL, Gabriella Vigliocco, UCL system. These concepts referred to properties of . A promising recent approach to the prob- objects in sets, including Boolean expressions of lem of the learning and representation of word- features (e.g., “red or circle”), context-sensitive meaning takes as its starting point the hypoth- concepts (e.g., “the largest in the set”), and con- esis that nontrivial aspects of the meaning of cepts requiring quantification (e.g., “every shape words can be inferred from their statistical dis- such that there is another shape of the same color tribution across spoken and written language. in the set”). We show that subjects are capable Numerous large-scale computational implemen- of easily inferring many of these concepts, often tations have successfully demonstrated the va- learning abstract, context-dependent rules that lidity of this so-called distributional hypothesis. involve explicit use of set-theoretic and quan- Important as these computational models have tificational operations. We then present a com- been, one of their widely shared practices has putational model which is similarly capable of been to treat the linguistic contexts in which learning these concepts and provides a good fit a word occurs as unordered sets of words. By to human learning curves. The computational disregarding all fine-grained sequential and syn- model formalizes probabilistic inference over a tactic information, however, these models dras- structured representation language for express- tically restrict the information from which word ing possible meanings. We compare the perfor-

3 mance of several representation languages in pre- word learning is best modeled as a process of dicting the human generalizations and show that statistical inference about speakers’ communica- people’s patterns of induction require a language tive intentions. In this talk I will describe a capable of set-theoretic, and quantificational op- Bayesian model of word learning based on this erations. This work extends classical rule-based idea. The model takes as its input coded corpus concept learning studies to representational sys- data from mother-child interactions and makes tems of sufficient formal power to represent com- joint inferences about what is being talked about positional linguistic meanings, and provides an (the mother’s intended referent) and the lexicon explicit computational theory for how children of the language being used. Making these joint may learn words that perform abstract semantic inferences allows the model to achieve consider- functions. ably higher performance than comparable asso- ciative models (which do not make any assump- Tuesday, 10:00 (Session A) tions about the speaker’s intentions) even with- One-Shot Learning with a Hierarchi- out including additional information. I will also cal Nonparametric Bayesian Model. describe ongoing work on including cues such as Ruslan Salakhutdinov, Joshua Tenen- eye-gaze and pointing into the model. In addi- baum, Antonio Torralba, MIT . In typical tion to performing better in learning from corpus applications of machine classification , data, our model allows us to make novel develop- learning curves are measured in tens, hundreds mental predictions about the phenomenon of dis- or thousands of training examples. For humans ambiguation (mutual exclusivity). We test some learners, however, the most interesting regime of these predictions with and children and occurs when the training data are very sparse. find that our data support a probabilistic view Just a single example of a novel category is often of mutual exclusivity but are not consistent with sufficient for people to grasp a concept and make either pragmatic or lexical principles approaches. meaningful generalizations to novel instances. In this talk we will present a nonparametric hierarchical Bayesian model that aims to capture Tuesday, 11:10 (Session A) this human-like pattern of one-shot learning. Using a developing lexicon to con- The proposed model leverages higher-order strain phonetic category acquisition. knowledge abstracted from previously learned Naomi Feldman, Brown University, Thomas categories to estimate the new category’s pro- Griffiths, University of California, Berkeley, totype as well as an appropriate similarity Sharon Goldwater, University of Edinburgh, metric from just one example. We provide an James Morgan, Brown University . Infants efficient MCMC and show on several begin segmenting and categorizing words as real-world datasets that the model is able to early as 6 months, mapping tokens heard in learn useful representations of novel categories isolation onto tokens of the same word heard based on a single training example. in fluent sentences, and this ability contin- Tuesday, 10:50 (Session A) ues to develop over the next several months Early word learning through communica- (Bortfeld et al., 2005; Jusczyk & Aslin, 1995; tive inference. Michael Frank, Stanford Jusczyk et al., 1999). During the same time University . How do children learn their first period, they show evidence of learning native words? While they are able to make use of dis- language phonetic categories (Werker & Tees, tributional information about the co-occurrence 1984). We use a hierarchical Bayesian model of words and objects, even very young chil- to explore how building a rudimentary lexicon dren also seem to take into account informa- can help a learner acquire phonetic categories tion about speakers’ communicative intentions. more robustly. The model learns to categorize Rather than being though of as purely statisti- speech sounds and words simultaneously from a cal or purely social, I argue that children’s early corpus of segmented acoustic tokens. No lexical

4 information is given to the model a priori; it weak ; and 3) these effects exist for re- is simply allowed to begin learning a set of call but not recognition. The second experiment word types at the same time that it learns to tested recovery interference by using a semantic categorize speech sounds. Simulations compare cue to train recovery of a competing response. As this model to a purely distributional learner predicted, this produced a deficit even that does not have feedback from a developing though the original target was not referenced by lexicon (McMurray et al., 2009; Vallabha et this training. al., 2007). The models are tested on a corpus Monday, 9:20 (Session A) constructed from CHILDES frequency data (Li A REM model of retrieval-induced for- & Shirai, 2000; MacWhinney, 2000) and English getting. Eddy Davelaar, Birkbeck, Univer- vowel production data (Hillenbrand et al., sity of London . The retrieval-induced forget- 1995), and results show that whereas a distri- ting paradigm consists of three phases. In phase butional learner mistakenly merges several sets 1, participants study a number of category- of overlapping categories, an interactive model exemplar pairs in sequence. Typically, each cat- successfully disambiguates these categories. egory label is paired six exemplars. In phase This provides one example of how feedback from 2, half of the presented category labels are pre- a developing lexicon can potentially improve sented as a cue together with some letters of learning in other domains, helping make the the target exemplar. The participants task is learning problem more tractable. to retrieve the target exemplar. However, only half of the exemplars that were paired with the Models of episodic memory. (Organizer: category label are retrieved, creating two ex- Simon Dennis) perimental conditions: non-retrieved items (Rp- ) and retrieved items (Rp+). In the third Monday, 9:00 (Session A) phase, all exemplars are probed and the mem- Learning to Forget: An Interference ory tested. Comparing the non-practiced items Account of Cue-Independent Forgetting. (Nrp) against Rp+ shows the basic learning ef- Tracy Tomlinson, University of Maryland, fect, whereas comparing them against Rp- shows College Park, David Huber, University of Cali- a disadvantage for the Rp- items. This be- fornia, San Diego . In memory inhibition theory, havioural inhibition has been discussed in the lit- forgetting is intentional, serving to eliminate un- erature as evidence for a mechanistic inhibition, wanted or distracting memories. The signature i.e., memories are permanently erased. The alter- of memory inhibition is cue-independent forget- native view is that the performance difference is ting in which a target memory is not recalled due to interference. This debate is played out us- even in response to an episodically unrelated se- ing recall data. The interesting point here is that mantic cue. In contrast, forgetting is due to in- the strongest argument fro an inhibition account terference in global matching memory models. is the presence of below-baseline performance in We demonstrate that these models can explain recognition memory. In this talk, I will present a cue-independent forgetting if there is interference REM-based model of retrieval-induced forgetting within the recovery stage of retrieval. A byprod- that captures the recognition data. The choice uct of learning to forget is that a sampled (not for REM in this simulation study is that REM yet recovered) memory is now associated with is an interference-based model. By showing its some alternative response, which explains why a capability of capturing retrieval-induced forget- deficit can exist regardless of the cues used to ting in item recognition, the view of mechanistic probe memory. We report 2 experiments that inhibition is further questioned. The simulation tested this account. The first experiment used study also shows the benefit of modelling in in- the think/no-think paradigm, revealing 1) sim- terpreting ”inhibition” findings. ilar forgetting when people are instead trained to press enter; 2) greater forgetting for initially Monday, 9:40 (Session A)

5 Using cued recall to move beyond task pressed as a function of internal time. Recent specific models. Amy Criss, William Aue, events are expressed with greater precision than Syracuse University . Free recall and single item less recent events. In addition to a variety of recognition have been extensively studied from timing behaviors, this model can account for re- an empirical and modeling framework. A num- cency and contiguity across time scales, as well ber of variables have different effects on perfor- as the activity of “time cells” observed in the ro- mance in these two tasks. These dissociations dent hippocampus. We discuss the potential for have inspired and informed theory development, applications to a broad variety of other episodic but they have also contributed to development of memory tasks, including judgments of recency models devoted to accounting for performance in and frequency. just one task and neglecting the other task. Cued recall shares properties with single item recogni- Monday, 10:50 (Session A) tion and free recall, making it a candidate task to The SARKAE model: Experience with resolve dissociations between the two tasks and novel characters determines recognition connect these two fields. The independent ef- memory. Pt I. Angela Nelson, UCSD, fects of word frequency and context variability , Indiana . Episodic mem- on cues and targets are evaluated in a series of ory must involve storage in and retrieval from experiments. Implications for current models of both general knowledge and event memory. Al- memory and the prospects of future models are though experiments typically control event pre- discussed. sentations, general knowledge is uncontrolled and highly variable across participants. We Monday, 10:00 (Session A) therefore trained knowledge for novel Chinese A distributed representation of remem- characters with training studies that continued bered time. Marc Howard, Karthik for several weeks. Of particular interest was Shankar, Syracuse . Episodic memory has been the role of character frequency: frequency var- described by verbal theorists as a form of “men- ied geometrically across characters during train- tal time travel” in which the rememberer jumps ing. Following this training, we gave tests of per- back in time to partially relive a previous event. ception, knowledge retrieval, and episodic recog- Distributed memory models (DMMs) have been nition memory. The recognition memory tests applied to numerous episodic memory tasks, in- proved highly diagnostic for delimiting models. cluding item recognition, cued recall and free re- These talks describe the data from several stud- call. We argue that while DMMs have provided ies, and the model they suggest for episodic a reasonable description of temporal effects in recognition memory. Study one used a visual episodic memory, including recency and contigu- search task for training. High frequency charac- ity, thus far they have not incorporated a suffi- ters tended to co-occur (as either targets or foils) ciently detailed model of history to provide a sat- across trials of training. The knowledge forma- isfactory description of episodic memory. For in- tion model we used stipulates that items thought stance, the temporal context model (TCM) mod- about together (e.g. because they occurred in els episodic memory as operations on a tempo- spatial/temporal proximity) tend to encode fea- ral context vector that changes during study and tures from each other, in laying down both event test. In TCM, the entire history with an item is traces and knowledge traces. Thus at the end thus expressed as a scalar value. Collapsing his- of training higher frequency characters became tory in this way prevents a solution of some ap- more similar to each other in the knowledge base. parently basic phenomena in learning. We sketch Large frequency effects were observed in testing a distributed representation of remembered time (and a mirror effect), but the roles of similar- that can be constructed from a set of tempo- ity and pure frequency were confounded in this ral context vectors with different time constants. study. Thus the second study trained knowledge In this model, the history with each item is ex- about the Chinese characters in a way that would

6 not alter inter-character similarity: Each char- nition memory. The recognition memory tests acter had to be matched to itself, for a decision proved highly diagnostic for delimiting models. whether a slight physical alteration had occurred. These talks describe the data from several stud- Again frequency effects were observed in the ies, and the model they suggest for episodic post-training recognition task. We were led to an recognition memory. Study one used a visual episodic recognition model in which summed fa- search task for training. High frequency charac- miliarity (summed event trace activation) is used ters tended to co-occur (as either targets or foils) to make a recognition decision. Performance is across trials of training. The knowledge forma- degraded by two sources of noise: activation of tion model we used stipulates that items thought traces of other characters in the study list, and about together (e.g. because they occurred in activation of training traces of the test item. In spatial/temporal proximity) tend to encode fea- study two (since similarity is equated across fre- tures from each other, in laying down both event quency) all frequency effects are produced by ac- traces and knowledge traces. Thus at the end tivation of training session traces. This model of training higher frequency characters became was tested in study two by a delayed recogni- more similar to each other in the knowledge base. tion test given after six weeks post-training. Fre- Large frequency effects were observed in testing quency effects tended to disappear, consistent (and a mirror effect), but the roles of similar- with the model. The model had one more critical ity and pure frequency were confounded in this feature: Because study two revealed strong and study. Thus the second study trained knowledge persistent effects of frequency for pseudo-lexical about the Chinese characters in a way that would decision, a role for pure frequency was required not alter inter-character similarity: Each char- and incorporated in the model: Encoding of any acter had to be matched to itself, for a decision presented character from knowledge was assumed whether a slight physical alteration had occurred. to be more accurate for higher frequency charac- Again frequency effects were observed in the ters. This factor has the effect of slightly im- post-training recognition task. We were led to an proving episodic recognition accuracy for higher episodic recognition model in which summed fa- frequency characters, but the model nonetheless miliarity (summed event trace activation) is used predicted both poorer overall recognition perfor- to make a recognition decision. Performance is mance and faster overall pseudo-lexical decision degraded by two sources of noise: activation of latencies for higher frequency characters. traces of other characters in the study list, and activation of training traces of the test item. In Monday, 11:10 (Session A) study two (since similarity is equated across fre- The SARKAE model: Experience with quency) all frequency effects are produced by ac- novel characters determines recognition tivation of training session traces. This model memory. Pt II. Richard Shiffrin, Indi- was tested in study two by a delayed recogni- ana, Angela Nelson, UCSD . Episodic mem- tion test given after six weeks post-training. Fre- ory must involve storage in and retrieval from quency effects tended to disappear, consistent both general knowledge and event memory. Al- with the model. The model had one more critical though experiments typically control event pre- feature: Because study two revealed strong and sentations, general knowledge is uncontrolled persistent effects of frequency for pseudo-lexical and highly variable across participants. We decision, a role for pure frequency was required therefore trained knowledge for novel Chinese and incorporated in the model: Encoding of any characters with training studies that continued presented character from knowledge was assumed for several weeks. Of particular interest was to be more accurate for higher frequency charac- the role of character frequency: frequency var- ters. This factor has the effect of slightly im- ied geometrically across characters during train- proving episodic recognition accuracy for higher ing. Following this training, we gave tests of per- frequency characters, but the model nonetheless ception, knowledge retrieval, and episodic recog-

7 predicted both poorer overall recognition perfor- To start to try to address the issue, we had mance and faster overall pseudo-lexical decision subjects wear a Microsoft SenseCamTMdevice latencies for higher frequency characters. that recorded images every 10-30 seconds over a period of a week. To understand the structure of Monday, 11:30 (Session A) their episodic experience, we analyzed the graph Extending the Context Maintenance and structure and dimensionality of these images. Retrieval model of free recall. Sean Polyn, We show that these episodic graphs have a Vanderbilt University, Michael Kahana, Uni- small-world structure which is characterized by versity of Pennsylvania . The context mainte- sparse connectivity, short average path lengths, nance and retrieval (CMR) model of memory and high global clustering coefficients. The search is a generalized version of the temporal conjecture is that the temporally contiguous context model of Howard and Kahana (2002). components of these clusters capture the ex- CMR proposes that memory search is driven by ternal component of the notion of context. A an internally maintained context representation degree distribution analysis shows that the composed of -related and source-related graphs are not scale-free, however, a correlation information. CMR suggests that organizational dimension analysis shows they are close to effects (the tendency for related items to cluster having a single dimensionality across scales. during the recall sequence) arise as a consequence of associations between active context elements and features of the studied material. The CMR Practical applications of models for re- model has been successful in accounting for orga- sponse time. (Organizer: Joachim Van- nizational phenomena, as well as serial position dekerckhove) effects and inter-response latencies, in both clas- Sunday, 9:00 (Session A) sic and new datasets (Polyn et al., 2009). An im- A cognitive psychometrical model for portant avenue of development is to bring CMR speeded semantic categorization decisions. into contact with other models of episodic mem- Joachim Vandekerckhove, Francis Tuer- ory. I will describe current work with the model linckx, Steven Verheyen, University of Leu- in which we examine the tension between the pre- ven . Choice reaction times (RTs) are often used dictions of context-based models such as CMR, as a proxy measure of typicality in semantic cat- and distinctiveness-based models such as SIM- egorization studies. However, other item proper- PLE (Brown et al., 2007). These recent inves- ties have been linked to RTs as well. We ap- tigations involve behavioral data from two new ply a tailored process model of choice RT to free recall studies, one examining the interaction a speeded semantic categorization task in order between category information and temporal in- to deconfound different sources of variability in formation, and the other examining a within-list RT. Different aspects of the response process are manipulation of temporal isolation (as in Brown then linked to different types of item properties. et al., 2006). Typicality measures turn out to predict the rate Monday, 11:50 (Session A) of information uptake, while lexicographic mea- What is context? Simon Dennis, Vishnu sures predict the stimulus encoding time. Avail- Sreekumar, Yuwen Zhuang, Mikhail ability measures cannot predict any component Belkin, Ohio State University . Many of us of the decision process. Applying the diffusion purport to study episodic memory, which by model to the data at hand required its extension definition is the ability to identify the occurrence to a hierarchical Bayesian framework, so that a of a stimulus in a specific context. Yet, outside “crossed random effects diffusion model” could the experimental setting, what constitutes a be constructed. This model retains the interest- context is an elusive question. If one proposes ing process interpretation of the diffusion model’s that recognition memory is dominated by con- parameters, but it can be applied to choice reac- text noise, then this difficulty is compounded. tion times even in the case where there are few

8 or no repeated of each participant- distribution increases as a function of target item combination. strength. They account for the empirical data by adopting a stricter criterion for strongly en- Sunday, 9:20 (Session A) coded lists relative to weakly encoded lists. The Modeling the Psychomotor Vigilance differentiation and criterion shift explanations Test for Evaluating Sleep Deprivation. have been difficult to discriminate using accu- Roger Ratcliff, The Ohio State University, racy measures. Reaction time distributions and Hans van Dongen, Washington State Univer- accuracy measures are collected in a list strength sity . I will present a model for the psychomo- paradigm and in a response bias paradigm where tor vigilance test. This task has become a stan- the proportion of test items that are targets was dard in assessing the behavioral effects of sleep manipulated. Diffusion model analyses are used deprivation. In the task, a millisecond timer re- to discriminate between these memory model ac- mains off for between 2 and 10 seconds and then counts. is turned on. The subject has to hit a button as soon as he/she detects that the counter is run- Sunday, 10:00 (Session A) ning. A critical finding is that when subjects are Application of the Ratcliff Diffusion Model sleep deprived, they show lapses in performance, to Reaction Times in a Two-Choice Au- that is, produce long reaction times (RTs). The ditory Detection Task. Siu Lung Lo, decision process used is a one boundary diffusion Nicholas Huang, Laurel Carney, Univer- process (inverse Gaussian or Wald distribution sity of Rochester, Fabio Idrobo, Boston Uni- of finishing times). This is preceded by a ran- versity . Abstract Amplitude modulations of dom Gaussian process that represents vigilance acoustic stimuli are essential for carrying infor- or readiness; it produces a delay prior to a signal mation in complex sounds and detection in noisy to turn on the decision process. The model fits environments. We are studying the relationship lapses (proportion of responses > 500 ms say) between and physiology in response and the RT distributions, including hazard func- to amplitude modulations. Our purpose was to tions. It fits these for sleep deprived and non- apply the Ratcliff Diffusion Model (RDM) to de- deprived sessions. The model parameters track scribe behavioral and physiological results in a alertness measures and they are correlated with study of detection of amplitude modulations of parameters from fits to data collected at the same either tone or noise stimuli. This study involved levels of deprivation from a two-choice numeros- a Bayesian tracking two-choice auditory detec- ity discrimination task. I will also discuss appli- tion task in which subjects (human and rabbit) cation of the model to simple RT paradigms and were instructed to discriminate between modu- relate the fits to two choice decisions. lated and unmodulated stimuli. Task difficulty Sunday, 9:40 (Session A) (condition) was manipulated by systematically Using the diffusion model to discriminate varying modulation in an effort to determine between memory models. Amy Criss, Syra- the detection threshold for amplitude modula- cuse University . In differentiation models, the tion. Traditionally, studies of detection using processes of encoding and retrieval produce an two-choice procedures report only threshold esti- increase in the distribution of memory strength mates, and occasionally also report d0 and bias. for targets and a decrease in the distribution of In this study the RDM was used to determine memory strength for foils as the amount of en- if these thresholds are percept driven, either by coding increases. This produces an increase in an accumulation of information or by other de- the hit rate and decrease in the false alarm rate cision factors. Various nested RDMs that dif- for a strongly encoded compared to a weakly en- fered in parameter constraints were used to fit coded list, consistent with empirical data. Other the reaction times using the DMAT toolbox. In models assume that the foil distribution is un- general, the RDM model provided a good fit to affected by encoding manipulations or the foil the reaction time distributions. Across a range of

9 modulation frequencies, results showed that drift Sunday, 11:10 (Session A) rate and probability of a modulated response Combining fMRI and the LBA model to increased with modulation depth, but they did ascertain a supramodal accumulator in not show evidence for a leading-edge effect, sug- perceptual decision making. Tiffany Ho, gesting the need to consider physiological mech- UCSD, Scott Brown, University of Newcas- anisms that could yield categorical decisions in tle, Australia, John Serences, UCSD . Reac- response to modulated sound. This RDM-based tion time models have been successfully utilized analysis enhances interpretation of the behav- in as tools for describing ioral results and provides constraints for models and analyzing behavioral data. Far less common of auditory processing. has been the use of RT models in brain data— specifically BOLD signals measured in fMRI. In Sunday, 10:50 (Session A) this talk, I will describe a study that illustrates Practice decomposed: stimulus specific the efficacy of a RT model of behavior in an- versus task specific effects. Gilles Dutilh, alyzing fMRI data. Here, we used the linear Eric-Jan Wagenmakers, University of Ams- ballistic accumulator (LBA) model (Brown and terdam, Joachim Vandekerckhove, Francis Heathcote, 2008) to ascertain supramodal deci- Tuerlinckx, Leuven University . When people sion making sites in the brain. Subjects par- repeatedly practice the same cognitive task, their ticipated in a perceptual decision making task response times (RTs) invariably decrease. The which comprised of varying levels of difficulty mathematical function that best describes this and which required two different modes of re- decrease has been the subject of intense theoret- sponse (saccades and button presses) while un- ical debate. Most theories of practice start from dergoing fMRI. Fits from our behavioral data the assumption that the practice effect is driven were used to make predictions about BOLD re- by changes in a single psychological process. In sponses for a supramodal evidence accumulator. our first study, we aimed to test this assumption Our analysis revealed that such an accumulator of a unitary process, and to take into account the exhibits bigger and temporally extended BOLD distribution of RTs and accuracy. Therefore, we responses characterized by a slower rise time on set out to decompose the effect of practice with the relatively difficult trials. Based on this analy- the Ratcliff diffusion model. A diffusion model sis, we were then able to identify the right insula analysis of data from a 10,000-trial lexical de- as a candidate supramodal accumulator. By us- cision task demonstrates that practice not only ing a combination of RT models and fMRI, the affects speed of information processing, but also mapping between model parameters (e.g., drift affects response caution, response bias, and pe- rate) and cognitive explanations (e.g., quality of ripheral processing time. In a second experiment, sensory information extracted) can be made all we aimed to further disentangle these effects into the richer with the addition of a neural basis. stimulus specific and task specific components. The data of this transfer experiment, in which Sunday, 11:30 (Session A) repeatedly presented sets and new sets of stimuli Developing and testing RT models for new were alternated, show that the practice effects on paradigms. Corey White, Roger Rat- speed of information processing are largely stim- cliff, The Ohio State University . We focus ulus specific. The effects of response caution and on one method to broaden the application of bias appear to be task specific. The effects on RT models, applying them to novel paradigms. peripheral processes, however, are partly stimu- Often this involves adjusting assumptions in lus specific and partly task specific. We conclude the original models and incorporating constructs that the diffusion model decomposition provides from domain-specific theories. We argue that any a perspective on practice that is more detailed new version of an RT model should satisfy three and more informative than the traditional anal- main criteria. First, any addition assumptions ysis of mean response times. or components that are added to the original

10 model should favor parsimony, both in number of models are used to estimate the cultural saliency parameters and psychological viability. Second, of each item and the cultural competence and the newer model must be shown to adequately fit response bias characteristics of each informant. the relevant behavioral data, and the new model Thus far there are CCT models for true/false, components should reflect the appropriate psy- multiple choice, digraph aggregation, matching chological constructs. Third, researchers must tests, and continuous response tasks. The talk demonstrate that the additional assumptions or will describe some of the available models and components do not adversely affect the original application areas of CCT, and contrast it to components of the model, and show that the other approaches to information aggregation. standard model components still track the ap- Sunday, 15:40 (Session A) propriate psychological constructs. We illustrate Hierarchical models for improving individ- these principles with a comparison of diffusion uals’ subjective probabilities. Ed Merkle, models of the flanker task and identify a simple Wichita State University . Common measures model that satisfies all three criteria, providing describing the “goodness” of subjective proba- support for its use as an analytical tool. bilities (e.g., those based on calibration and dis- crimination) fail to capture the full extent to Cognitive modeling and the wisdom of which subjective probability and accuracy are re- crowds. (Organizer: Mark Steyvers) lated. A judge may exhibit poor calibration and discrimination, yet still report subjective prob- Sunday, 15:20 (Session A) abilities that are predictive of accuracy. In the Cultural Consensus Theory. talk, I describe hierarchical models designed to William Batchelder, UCI . Cultural exploit these relationships, transforming individ- consensus theory (CCT) is a formally based uals’ subjective probabilities without aggregat- methodology for determining consensus answers ing them. The resulting, transformed probabil- to questions from groups of informants who ities exhibit improved statistical properties that share knowledge or beliefs. It was developed can also improve aggregation across individuals. in the 1980s by myself, A. Kimball Romney, Specific models studied within this framework in- and colleagues, e.g. Batchelder & Romney clude hierarchical logistic regression models and (1988, Psychometrika); Romney, Weller, & error models of confidence. The partial pooling Batchelder (1986, American Anthropologist), inherent in the models implies that the crowd and since then CCT has become the leading influences the way in which individuals’ proba- methodology for determining consensus beliefs bilities are transformed. within the areas of cognitive anthropology, e.g. Weller (2007, Methods). CCT consists of Sunday, 16:00 (Session A) a collection of cognitively motivated, parametric Wisdom of the Crowd Within: Sampling statistical models for aggregating the responses in Human Cognition. Ed Vul, MIT . Across of informants to a series of test items about some a number of tasks, people approximate optimal domain of their shared knowledge or beliefs. It Bayesian behavior, but exact Bayesian compu- is based on a testable assumption that there tations are impossible to carry out – how can are consensus (culturally correct) answers to the cognition approximate ideal Bayesian inference items based on the common culture of the in- despite limited cognitive resources? I argue that formants; however, there is no requirement that human cognition approximates Bayesian infer- the answers correspond to truths of the natural ence by sampling. A prediction and consequence world. The models generally take the form of approximate inference by sampling is indepen- of models in psychometric test theory except dent, identically distributed noise across multi- the correct answers are specified as parameters ple responses from one individual. Here I will in the models rather than known apriori. In review several experiments yielding support for addition to estimating consensus answers, the this claim. I show that in visual attention tasks

11 (selection in space and time) multiple guesses the judgment of percentages and rankings of about the identity of the target are independent events and items. We compare situations in and identically distributed. Further, in world- which individuals independently answer these trivia tasks, when subjects are asked to make two questions with two group-decision making guesses about questions like “What proportion of situations: an iterated learning environment in the world’s airports are in the US?”, their guesses which individuals pass their solution to the next are also independent. As a consequence of this person in a chain and a Delphi-like method in independence, averaging two guesses from one in- which individuals can revise their answers after dividual yields a wisdom of a “crowd within”: observing the solutions from other members of averaging multiple guesses from one individual the group. We extract the shared knowledge in yields a more accurate answer than either of the all these situations using a hierarchical Bayesian guesses alone. approach inspired by Cultural Consensus The- ory. Consistent with previous research, the Sunday, 16:20 (Session A) results indicate that group decision-making sit- Representing consensus and divergence in uations lead to correlated judgments. However, communities of Bayesian decision-makers. we show that a proper analysis of the data Kshanti Greene, University of New Mexico . that takes the dependencies between individuals We address the concept of wisdom of the crowds into account leads to strong wisdom of crowd for subjective problems such as policy-making in effects where the aggregate performs as well as which the population will typically have very di- or better than the best individual in the group. verging opinions and there is no single “correct” answer. We introduce a new approach to ag- gregating the beliefs and preferences of groups of individuals and demonstrate how to apply ACCEPTED TALKS game theoretic analysis to large populations of Sunday, 10:50 (Session C) Bayesian decision-makers. Belief aggregation ap- Subjective Loudness As A Ratio Scale proaches that form a single consensus model av- Function of Both Intensity And Frequency. erage away conflicting beliefs in a population and R. Duncan Luce, Ragnar Steingrimsson, may even result in irrational solutions according Louis Narens, UCI . Using the methods of to a set of properties for rational aggregation. In magnitude estimation and production, S. S. contrast, our approach maintains the diversity of Stevens (1975) claimed without proof that when a population and allows the competitive aspects all other physical properties other than inten- of group decision-making to emerge. We form sity are held fixed, then subjective intensity is collective belief models in which collectives are a ratio scale of physical intensity. Narens (1996) formed from individuals who agree on the pref- proved that for this to be true a certain commu- erence order over a set of decision alternatives. tativity property must hold, and several empiri- Using “super-agents” that represent each collec- cal studies confirmed commutativity for loudness tive we demonstrate how game theoretic analysis and for brightness. We address a more general and negotiation techniques can be used to enable question: Is the loudness function over all inten- and analyze rational social decision-making. sities and frequencies a single ratio scale that, for Sunday, 16:40 (Session A) each frequency separately, is the usual loudness A Bayesian approach to aggregate knowl- scale? Using Luce’s (2004) axiomatic represen- edge: Comparing groups of indepen- tation, which is more general than Narens’, the dent and communicating individuals. answer is yes provided that a more general com- Mark Steyvers, Brent Miller, UCI , Rob mutativity property holds. Data in support of Goldstone, Indiana University . We analyze generalized commutativity are reported. A ques- the collective performance of individuals in tion to be addressed in the future is: If another a series of general knowledge tasks involving modality is also represented by a ratio scale, is

12 there a common subjective intensity over the two ple pieces of furniture. The model consists of sev- modalities? eral steps. First, individual objects are found in the 2D image and separated from each other and Tuesday, 16:20 (Session C) from the background. The image is obtained by Emergence of Cooperation between Intel- a calibrated camera (the focal distance and the ligent Individuals with Personality. Min- principal point are known). This is accomplished Hyung Cho, Soo-Young Lee, KAIST .In by a combination of region growing and contour this paper, we have shown that cooperative be- detection. The occluding contour of a single ob- havior can be spontaneously evolved among intel- ject is represented by an alpha-hull of a region en- ligent individuals by adopting personality. Evo- compassing homogeneous texture. Alpha-hull is lutionary prisoner’s dilemma game in a grey-scale a generalization of a convex hull. Next, contours version, which we use as a model of real society, inside each alpha-hull are detected and grouped. abstracts the essence of interaction between liv- These internal contours are essential for produc- ing beings. Our model of for play- ing a 3D interpretation of each object. The 3D ers is composed of prediction network and ac- shape of the object is recovered by applying sev- tion network. The prediction network models the eral a priori constraints (3D symmetry, orthogo- environment based on the information from the nality, 2D and 3D compactness, planarity of the outside. The action network simulates N-look contours). An inclinometer mounted on the cam- ahead future and choose the next action that era provides information about the direction of is expected to induce the best future outcome. gravity. This information is similar to the in- The first proposed model can exploit or coop- formation provided by the human vestibular sys- erate with fixed strategy players without diffi- tem. Information about gravity is used to deter- culty, thus guaranteeing a moderate intellectual mine the horizon and the vanishing point repre- power. But it was hard to induce cooperation senting edges that are vertical in the 3D scene. between proposed models because they could not The 3D scene is recovered by placing the individ- get the accurate model of others and easily fell ual 3D shapes on the common floor, whose dis- into mutual defection with no investment. By tance from the camera is known. The recovered adopting personality to the first model the sec- scene is usually very close to the real one and it ond model has solved this problem. We have is complete in the that it has information tried two different characterizations of person- about the back, invisible parts of the objects as ality, namely CO and RS, and evolution experi- well as the front, visible ones. ments have yielded different dominant intelligent strategies. For CO characterization case, which Sunday, 16:40 (Session B) compromises selfishness and altruism, the usual Any 2D image is consistent with mutual defection strategy has become dominant a 3D symmetrical interpretation. out of initial random population. Whereas, for Tadamasa Sawada, Yunfeng Li, Zyg- RS characterization which represents the level of munt Pizlo, Purdue University . Our everyday greediness, mutual cooperation strategy has be- life experience suggests that we can easily come dominant as a result of evolution. Two final tell the difference between 3D symmetrical strategies can also be shown to be evolutionarily and asymmetrical shapes. Computationally, stable. discriminating between 3D symmetrical and Sunday, 16:20 (Session B) asymmetrical shapes from a single 2D retinal A computational model of the recovery of image is a difficult problem because the 2D a 3D scene from a single 2D camera image. retinal image of a 3D shape is almost always Yunfeng Li, Joseph Catrambone, Stephen asymmetrical regardless whether the 3D shape Sebastian, Tadamasa Sawada, Zygmunt Pi- is or is not symmetrical. One possible method to zlo, Purdue University . Our recovery model is perform this discrimination is to verify whether applied to images of real scenes containing multi- a given 2D retinal image is geometrically con-

13 sistent with a 3D symmetrical interpretation. An Integrated System Analysis of Gain Unfortunately, this method will not work and Orienting Models of Attentional Se- because 3D symmetrical interpretations are lection. Philip Smith, David Sewell, The almost always possible. We prove, under very University of Melbourne . The classical attention general assumptions, that any pair of curves in literature offers two alternative conceptions of a single 2D image is consistent with a pair of the brain’s limited capacity processing resources 3D symmetrical curves. It is also trivially true and how resources are allocated: orienting mod- that any pair of curves in a single 2D image els and gain models. Orienting models have their is consistent with a pair of 3D asymmetrical origins in Broadbent’s filter theory and Posner’s curves. It follows, that discrimination between spotlight of attention. They hold that, in order 3D symmetrical and asymmetrical interpreta- to identify a stimulus, a limited-capacity, pos- tions is geometrically impossible. So, how can a sibly serial, stimulus identification mechanism human observer discriminate between symmet- must be switched to, or aligned, with the stim- rical and asymmetrical 3D shapes? We discuss ulus location. Gain models have their origins in the role of higher-order features (like corners) in Kahneman’s capacity theory. They hold that ca- determining the point correspondences, as well pacity can be distributed variably across stimu- as the role of a planarity constraint in this task. lus locations in accordance with attentional set, When the correspondence of features is known and that the processing efficiency at any location and/or curves can be assumed to be planar, is proportional to the capacity allocated there. 3D symmetry becomes non-accidental. In other Smith and Ratcliff (, 2009) words, a 2D image of a 3D asymmetrical shape compared gain and orienting models in their in- obtained from a random viewing direction is tegrated system theory of visual attention and unlikely to be consistent with 3D symmetrical found that their predictions in a Posner spatial interpretations. cuing task were indistinguishable. We report two new experiments based on Yantis and Jonides’ Monday, 16:40 (Session A) no-onset paradigm, in which stimuli are created Cognitive Models, the Wisdom of Crowds, by removing features from existing perceptual and Sports Predictions. Michael Lee, UCI objects. RTs to unattended stimuli were longer . The Wisdom of Crowds framework has been and accuracy was lower in a no-onset condition applied not just to general knowledge problems, than in a condition with onset transients. Fits of where there is an existing ground truth, but also the integrated system model show the data were to prediction problems, where the ground truth better described by a model in which process- is only established later. This talk presents some ing resources are reallocated in response to onset case studies in predicting sporting outcomes, ex- transients than one in which resources are allo- ploring the role that cognitive models can play cated only once, at the beginning of a trial. We to improve the aggregation of expert and non- characterize the results using Sperling and We- expert human opinions. One involves ichselgartner’s episodic theory, which views gain predicting NBA ladders, based on a mixture of and orienting as the spatial and temporal dimen- expert and non-expert predictions. Another in- sions of an attentional control function. volves predicting NFL game winners, based on regularly revised expert assessment. A third in- Tuesday, 15:20 (Session C) volves predicting the “March Madness” College A non-probabilistic measure of relational Basketball tournament, where there is strong information based on Categorical Invari- prior information. Collectively, these examples ance Theory. Ronaldo Vigo, Ohio Univer- highlight the potential and challenges in using sity . In what follows, we introduce a measure of hierarchical generative cognitive models to make relational information that is not based on prob- aggregated predictions. ability theory. The measure is based on a unifi- Tuesday, 15:20 (Session A) cation of invariance, complexity and information

14 notions in concept learning research. The idea is ROC analyses of confidence responses and to use categorical invariance theory (Vigo, 2008, reaction times in recognition memory. 2009) to characterize the subjective relational in- Christoph Weidemann, Swansea University, formation that a Boolean set R carries about its Michael Kahana, University of Pennsylvania superset S as the rate of change in the subjective . A central assumption in models of recogni- structural complexity of S (as defined by the de- tion memory is that memory strength is vari- gree of categorical invariance of S) whenever the able. In an effort to assess memory strength objects in R are removed from the set S. We ran on a trial-by-trial basis, many recognition stud- an experiment which shows that humans find an ies ask participants to qualify the binary recog- object more informative in respect to its cate- nition response with confidence ratings. We in- gory of origin when there is a relatively large de- vestigated the question to what extent informa- crease or large gain in the subjective structural tion contained in such confidence ratings might complexity of the category without the object; already be present in reaction times of binary likewise, less informative when there is a rela- “old”/“new” decisions. Novel ROC-based mea- tively small decrease or small gain in the subjec- sures quantify to what extent either dependent tive structural complexity of the category with- variable contains information about intermedi- out the object. Our subjective relational infor- ate states within each recognition response. We mation model successfully fits this data. This show that in many cases there is a close corre- talk also serves as an introduction to categori- spondence between information gained from con- cal invariance theory which posits that the de- fidence ratings and from reaction times to binary gree of subjective learning difficulty of a Boolean recognition decisions. category (and of its corresponding concept) is a function of its degree of categorical invariance. Monday, 10:50 (Session B) Computer-Based Assessment of Work- Sunday, 10:50 (Session B) ing Memory Using Computer Games. Towards a unified account of deci- Misha Pavel, Holly Jimison, OHSU .Asa sions from experience and description. part of a large longitudinal study focused on Scott Brown, Guy Hawkins, Pennie early detection of cognitive decline in the ag- Dodds, Ben Newell, Adrian Camilleri, ing population, we have been investigating ways University of Newcastle, Australia . Research on to monitor cognitive processes using unobtru- risky decision making has focused on problems sive techniques combined with real-life cognitive in which the choice alternatives are described tasks. We developed a number of computer to observers (“Would you prefer a sure gain of games which, much like in everyday cognitive $3 or an 80% chance of a $4 gain with a 20% tasks, require concurrent execution of multiple chance of no gain?”). Some more recent research cognitive processes. Mathematical models are re- has found that different decisions are often quired to estimate the key characteristics of the made observers learn these task parameters by underlying cognitive processes. In this presenta- repeatedly sampling outcomes from the alter- tion we illustrate this approach using a computer natives. We present a theoretical framework game of concentration. At the start, a player is that is aimed towards accounting for choice confronted with a matrix of images of playing preferences in both experience and description cards displayed face-down and asked to turn two paradigms. We also evaluate the model using a cards on each trial. The players goal on each new stream of data provided by the experience trial is to identify (i.e., to turn) two matching paradigm: observers not only provide choices cards using the player’s memory of previously between alternative gambles, but also judgments exposed card face value and location. The key about the outcomes and probabilities associated cognitive process required to play this game is with each gamble. working memory. A player’s working memory is Tuesday, 9:20 (Session B) represented by a leaky buffer of previously seen

15 cards; each item can be lost stochastically from cision. Louis Narens, UCI . Normative and the buffer. The duration of an item in working descriptive theories of belief and decision are al- memory is described in terms of survival analysis most always formulated in terms of a finitely characterized by a hazard function derived from a or countably additive probability function on a Weibull distribution. After removing the effects of sets. This talk investigates of guessing (i.e., a model of a random player), the looking at a weaker algebra of subsets upon parameters of the maximum likelihood fit of the which to base finitely and countably probability distribution are used to characterize the working theory for representing psychological and norma- memory of each player. We will demonstrate the tive representations of belief and decision. Philo- applicability of the model and its use to track sophical and modeling justifications for the new individual players’ working memory over time. are presented and applications to are discussed. Tuesday, 12:10 (Session B) Multinomial Processing Tree Models and Sunday, 9:20 (Session C) Bayesian Networks: Some formal results. Quantitative Testing of Decision Theories: Brendan Purdy, Moorpark College, William II. Empirical Testing. Chris Zwilling, Batchelder, University of California, Irvine . Michel Regenwetter, William Messner, Multinomial Processing Tree (MPT) Models are Anna Popova, University of Illinois . We carry class of statistical models applied to cognitive out a series of quantitative tests of leading theo- phenomena. They have been widely used within ries of decision making, including Expected Util- the field of mathematical psychology. It is im- ity theory, Cumulative Prospect theory and the portant to note that the tree structure of MPT Transfer-of-Attention-Exchange model. All of models is intended to represent cognitive pro- these tests avoid aggregation paradoxes by ana- cesses. The first set of propositions of this paper lyzing data within respondent only. The tests are concerns the enumeration and statistical equiva- based on novel “distance-based,” “aggregation- lence of various subclasses of MPT models. That based,” and “mixture-based” probabilistic spec- is, the paper states a number of results regard- ifications, all of which require state-of-the-art ing how many models there are for certain types order-constrained inference. All of these tests of MPT models and when these models can be turn out to be extraordinarily powerful. We re- said to be statistically equivalent to one another. inforce other authors’ warning that probabilis- The proofs of these propositions are built on the tic specification of algebraic theories cannot be formalization of Binary MPT models in Purdy taken lightly. and Batchelder (2009, JMP). These formal re- Monday, 16:00 (Session B) sults lead to a theorem in the paper about the Application of a Robust Differencing Vari- relationship between MPT models and Bayesian able (RDV) Technique to the Depart- networks. Bayesian networks (Bayes nets) are a ment of Veterans Affairs Learners’ Per- class of statistical models that are widely used in ceptions Survey. Richard Golden, Univer- both and cognitive science. In sity of Texas at Dallas, Steven Henley, Mar- particular, the theorem demonstrates that MPT tingale Research Corporation, Halbert White, models are a subclass of Bayes nets. The proof of University of California at San Diego, Michael this theorem depends on expanding the numbers Kashner, UT Southwestern Medical Center at of random variables that are built into the class Dallas . An important problem in policy anal- of MPT models. An important aspect of this pa- ysis is measuring the effect of a factor using per is the discussion of why these theorems are observational data by comparing response out- relevant to the researcher in the field or the lab. comes from a a “factor-present” condition with responses to a non-randomized control condition Sunday, 11:50 (Session C) where the factor is absent. We thus introduce A new probability theory for belief and de- the Robust Differencing Variable (RDV) Tech-

16 nique that applies Robust Missing Data The- grammar, and further the class of strings so ory (Golden, Henley, White, and Kashner, sub- defined coupled with the computational rules mitted). To isolate group factor effect size, a is shown to be formally equivalent to the class “differencing variable” is introduced to indicate of BMPT models defined in the traditional whether the factor, when present, will produce an way. This paper will extend the string language effect or produce no effect, on the response vari- in three directions. First we will generalize able. By their nature, differencing variables are the string language to handle multi-link MPT observable in the factor-present condition, but (MMPT) models where non- terminal nodes are usually unobservable in control conditions. can have more than two children. Second we Treating the partially or completely unobserv- will exhibit transformational rules on multi-link able differencing variable in the control condi- MPT model strings that generate statistically tion as Missing at Random (MAR) we show how equivalent BMPT model strings. Finally we to measure a group factor effect-size. This is will reconsider the problem of parametric order accomplished by computing the association be- constrains in BMPT models discussed in Knapp tween the response outcome and an interaction and Batchelder (2004, JMP) by developing between the differencing and group factor indica- string transformational rules that will generate tor variables so that the resulting maximum like- statistically equivalent, unconstrained BMPT lihood estimates, standard errors, and hypothe- models from strings representing BMPT models sis tests are robust to model misspecification. We subject to parametric order constrains. Viewed apply this strategy to the Department of Veter- generally the string language provides a very ans Affairs (VA) Learners’ Survey of succinct representation of BMPT models that medical physician residents to assess the impact can aid in theorem proving and facilitate input of an 80-hour work week limit on medical educa- to computational statistical algorithms. tion. Our results provided evidence that the 2003 clinical duty limits had positive impacts on clin- Monday, 15:20 (Session C) ical learning environments, a finding that offers Evolutionary models of color catego- profound policy implications on how physicians rization based on realistic observer will be trained in the U.S. in the future. models and population heterogeneity. Kimberly A. Jameson, Natalia Komarova, Tuesday, 11:50 (Session B) UCI . The evolution of population color cate- A String Language for Multi- gorization systems is investigated using game nomial Processing Tree Models. theory simulation techniques. Categorization William Batchelder, UC Irvine, Bren- systems are based on human empirical data dan Purdy, Moorpark College, Xiangen Hu, in the form of Farnsworth-Munsell 100 Hue U Memphis . Traditionally Binary Multinomial Test results to, in part, model artificial agents Processing Tree (BMPT) models are defined as with some realistic constraints from human rooted full binary trees, where manifest cate- populations. Constraints include (i) varying gories are associated with the leaves (terminal amounts of normal observer heterogeneity, and nodes) and parameters with the non-terminal (ii) varying degrees and forms of observer color nodes of the tree. Purdy and Batchelder deficiency, and are made operational in agent (2009, JMP) provide five recursive axioms categorization and communication games. They that provide a new definition of the class of produce a number of interesting consequences BMPT models. Those axioms represent BMPT for stable, shared categorization solutions that models as formal strings of parameter and are evolved in agent populations. It is found category symbols along with computational that the confusion patterns associated with a rules that apply to the strings to generate small fraction of color deficient agents break category probabilities. The string language symmetries in population categorization so- is shown to derive from a formal context free lutions, and confine the boundaries of color

17 categories to a small subset of available loca- the laboratory. We discuss several probabilistic tions. Further, confusion pattern variations specifications that all leverage breakthrough across different types of deficient agents lead developments in statistical inference. We also to changes in category size and number that illustrate the formal implementation of these depend on the form of deficiency represented. specifications for leading current theories of In particular, stimulus pairs forming global decision making. confusion axes for dichromats tend to attract color boundaries, and local confusion regions Sunday, 15:40 (Session B) characteristic of both dichromats and anomalous Human four-dimensional spa- trichromats tend to repel color boundaries. tial judgments on hyper-volume. Furthermore, the concurrent presence of normal Ranxiao Frances Wang, Univ. of Illi- agents and several different kinds of deficient nois at Urbana-Champaign . Living in a 3D agents produces novel constrained solutions that physical world, humans have their perceptual optimize successful categorization performance and cognitive systems tailored for 3D objects. in population communication games involving However, a recent study by Ambinder et al color. (2009) demonstrated humans ability to learn and make basic spatial judgments (distance Tuesday, 12:10 (Session C) and angles) on 4D geometric objects viewed in Remarks on quantum cognition. virtual reality with minimal exposure to the task Geoff Iverson, UCI . In October 2009 a and no feedback to their responses. The current special issue, Volume 53 (5), of the Journal case study further examines human judgments of Mathematical Psychology was devoted to of hyper-volume, a novel property unique to the emerging field of quantum cognition. In higher dimensional space. One experienced their introductory article the editors of this observer was tested in 20 trials. In each trial, special issue noted that “Several phenomena the participant studied a 3D parallel projec- in psychology that have stubbornly resisted tion of a random wireframe hyper-tetrahedron traditional modeling techniques are showing (4-simplexes) rotating along the wx-plane for promise with a new modeling approach inspired 2 minutes. Then the hyper-tetrahedron dis- by the formalism of quantum mechanics.” In appeared and a cube appeared representing examining this provocative claim I pay close at- the base of a hypercube. The participant tention to the way in which quantum formalism adjusted the size of the cube until the hyper- is implemented in the construction of cognitive volume of the hypercube it represented and models, and to the data that these models seek the hyper-tetrahedron matched. To account to explain. for the 3D strategy which uses averaged 3D Sunday, 9:00 (Session C) volume to approximate the hyper-volume, a Quantitative Testing of Decision The- multiple regression was run on the responses ories: I. Probabilistic Specification. as a function of both the actual hyper-volume Michel Regenwetter, University of Illinois, and the mean 3D volume to measure people’s Clintin Davis-Stober, University of Mis- true hyper-volume judgments. The correlation souri, Shiau Hong Lim, Universitaet Leoben, between the hyper-volume and the mean 3D vol- William Messner, University of Illinois . ume was moderate (r = .37), suggesting the two We provide a new quantitative framework for factors can be separated. More importantly, the testing algebraic theories of pairwise preference multiple regression showed a significant effect of on binary choice data. This framework bridges hyper-volume (β = .71, t =3.71, p<.01) but the conceptual, mathematical, and statistical not mean 3D volume (β = .04, t = .22, p = .82). gap between algebraic in the de- These findings suggest that at least some people terministic realm and highly variable empirical are able to judge 4-D hyper-volume, which data that originate from sampling processes in indicates the true 4D spatial representations.

18 Tuesday, 9:00 (Session B) tute . When people make decisions they often On Incorporating Item Effects into Signal face opposing demands for response speed and . Lawrence DeCarlo, response accuracy, a process likely mediated by Columbia University . Applications of signal de- response thresholds. According to the striatal tection theory (SDT) often involve the presen- hypothesis, people decrease response thresholds tation of different items on each trial, such as by increasing activation from cortex to striatum, different words in a memory study or differ- releasing the brain from inhibition. According ent slides in a medical imaging study. It has to the STN hypothesis, people decrease response long been recognized that factors particular to thresholds by decreasing activation from cortex the items themselves might affect observers re- to subthalamic nucleus (STN); a decrease in STN sponses. Here it is noted that there are differ- activity is likewise thought to release the brain ent ways to incorporate item effects into SDT, from inhibition and result in responses that are depending on how the situation is conceptual- fast but error-prone. To test these hypotheses— ized. For example, a latent continuous variable both of which may be true—we conducted two that represents item effects can be assumed to experiments on perceptual decision making in affect only one event, such as the signal, or can which we used cues to vary the demands for be assumed to affect both events, that is, sig- speed vs. accuracy. In both experiments, behav- nal and noise. An example is a medical training ioral data and analyses con- study where trainees attempted to detect frac- firmed that instruction from the cue selectively tures in X-ray slides of ankles (where the true affected the setting of response thresholds. In the state of the ankle was known). The SDT model first experiment, we used ultra high-resolution gives estimates of a trainees ability to discrimi- 7T structural magnetic resonance imaging (MRI) nate fractures from non-fractures and his or her to locate the STN precisely. We then used 3T use of response criteria. The idea underlying structural MRI and probabilistic tractography the extended SDT model is that the slides also to quantify the connectivity between the rele- differ in terms of difficulty, which can be cap- vant brain areas. The results showed that par- tured by a latent difficulty variable that affects ticipants who flexibly change response thresh- the discrimination parameter. A model where olds (as quantified by the mathematical model) only fractures differ in detection difficulty is com- have strong structural connections between pre- pared to a model where both fractures and non- supplementary motor area and striatum. This fractures differ in difficulty. The extended model result was confirmed in an independent second is a type of random coefficient model and can experiment. In general, these findings show that easily be fit with maximum likelihood estimation individual differences in elementary cognitive using standard software. The extension incorpo- tasks are partly driven by structural differences rates an item-response theory (IRT) model into in brain connectivity. Specifically, these findings an SDT model, with simultaneous estimation of support a cortico-striatal control account of how both components. the brain implements adaptive switches between cautious and risky behavior. Tuesday, 11:30 (Session C) Cortico-Striatal Connections Predict Con- Monday, 10:00 (Session B) trol over Speed and Accuracy in Percep- Looking to Learn, Learning to Look: At- tual Decision Making. Birte Forstmann, tention Emerges from Cost Sensitive In- University of Amsterdam, Alfred Anwan- formation Sampling. Bradley Love, Uni- der, Andreas Schafer¨ , Jane Neumann, Max versity of Texas at Austin . Given people’s ca- Planck Institute, Scott Brown, University of pacity limitations, one key aspect of learning is Newcastle, Eric-Jan Wagenmakers, Univer- learning which stimulus aspects are goal relevant sity of Amsterdam, Rafal Bogacz, University in the current context. In addition to injecting of Bristol, Robert Turner, Max Planck Insti- noise into the decision process, gathering unnec-

19 essary information can have costs in terms of the models agree on the effects of changes in the time, effort, dollars, fuel, etc. Accordingly, many rate of information accumulation and changes in category learning models employ selective atten- non-decision time, but that they disagree on the tion mechanisms that learn which stimulus di- effects of changes in response caution. Fits to em- mensions are most critical to performance. How- pirical data show, however, that the models agree ever, attention in category learning models does closely on the effects of experimental manipula- not direct what is encoded, but instead estab- tions, both within and across groups, even when lishes decision weights on stimulus dimensions. that manipulation affects response caution. We According to existing models, all stimulus dimen- conclude that—in real data—researchers can an- sions are encoded, which is not realistic, nor sup- alyze their data with either model, as the conclu- ported by basic research in attention. To address sions about psychological processes are unlikely these shortcomings, I develop a model that se- to depend on the model that is used. As further lectively encodes information during learning as confirmation of the interchangeability of the two a function of the learner’s goals, task demands, models we also find that they have equivalent and knowledge state. The model consists of two complexity. components that are both normative, but lead to Tuesday, 15:40 (Session A) apparent non-normative behaviors when linked. Guided Visual Search. Chris Donkin, Indi- One component determines the value of poten- ana University, Denis Cousineau, University tial sources of information. The value of a piece of Montreal, Richard Shiffrin, Indiana Uni- of information depends on the decision maker’s versity . Cousineau and Shiffrin (2004) presented goals and assumptions about (i.e., knowledge of) a model for the control conditions of a visual the world, as well as the cost of the information. search study. Multiple modes were observed for The second component of the model reflects the target present trials and the model posited se- decision maker’s knowledge of the world, which rial terminating comparisons, partially and prob- is used by the first component to direct infor- abilistically guided to the target location by a mation gathering. This learning component of separate parallel automatic process. We expand the model is updated by the information samples and extend this model to the experimental condi- selected by the first component, completing the tions in which the visual display objects are pre- cycle of mutual influence. The model accounts sented successively, at speeds fast enough that for accuracy and eye movement data during cat- the displays appear simultaneous. These succes- egory learning tasks. sive display conditions place strong constraints on models, and allow us to develop a quite pre- Sunday, 12:10 (Session A) cise account for parallel and serial processes op- Diffusion versus Linear Ballistic Accumu- erating together in visual search. lation: Different Models for Response Time, Same Conclusions about Psycholog- Monday, 16:00 (Session C) ical Mechanisms? Chris Donkin, Indiana Color charts, aesthetics, and subjec- University, Scott Brown, Andrew Heath- tive randomness. Yasmine Sanderson, cote, The University of Newcastle, Eric- Friedrich-Alexander University Erlangen . Much Jan Wagenmakers, University of Amsterdam, of the previous research in subjective randomness Richard Shiffrin, Indiana University . Quan- has considered it within the framework of binary titative models for response time and accuracy sequences and grids (see Bar-Hillel and Wage- are increasingly used as tools to draw conclu- naar (1991) or Nickerson (2002) for a survey). sions about psychological processes. Here we in- In this study, we consider subjective random- vestigate the extent to which these substantive ness within the additional framework of color. A conclusions depend on whether researchers use statistical analysis of fifty-four “random-looking” the Ratcliff diffusion model or the Linear Ballis- art and design color charts show that they differ tic Accumulator model. Simulations show that significantly from normal in the average jump be-

20 tween adjacent colors. The charts in question are Andrew Richardson, Dare Institute . This human-generated and show no obvious patterns study analyzed items from five variations of a or symmetries and are not likenesses of some per- problem that required the isolation of variables. son, place or object. An example of one such The purpose of this analysis was to determine chart is: the relationship between order of hierarchical complexity of items, their content, and language and participant performance on the items. Participants were English, German and Arabic speaking. The five variations of the isolation of variables problem that were given were: 1) Ara- Figure 1: Happy Pixels by Sieger Design bic Laundry; 2) United States Laundry; 3) Short Laundry; 4) Combustion; 5) Atheism and Belief. We argue that this (and many of the other) color A Rasch analysis produced stage scores for charts are examples of subjective randomness. each of the items from each of the instruments. Our results suggest that color distance is a factor These item Rasch scores were regressed against in subjective complexity. We also argue that the the order of hierarchical complexity of each of large percentage of design color charts with su- the items, r = .95 (Arabic Laundry); r = .963 peralternation indicate an aesthetic value to this (United States Laundry); r = .940 (Short trait. Laundry); r = .797 (Combustion); r = .787 (Atheism and Belief). The analysis showed Sunday, 9:00 (Session B) that Hierarchical Complexity was the greatest Adaptive Design Optimization for predictor of variance r = .898, p<.01 when Testing Generalized Utility Theories. compared against both content and language, Daniel Cavagnaro, Jay Myung, Mark together and individually. We showed that Pitt, Ohio State University, Richard Gon- the ordinal nature of the order of hierarchical zalez, University of Michigan . We present a complexity was reflected in gaps between stage Bayesian framework for comparing generalized scores. utility theories. In this framework, each theory is cast as a parametric Bayesian model based on Tuesday, 9:20 (Session C) a qualitative pattern of indifference curves (e.g., Can Perceived Value Be Explained parallel straight lines, straight lines fanning out, by Schedules of Reinforcement?. or convex curves). Bayesian model selection Nicholas Commons-Miller, Tufts Univer- tools can then be used to weigh the evidence for sity, Michael Lamport Commons, Harvard each theory. Data that can effectively discrimi- Medical School, Robin Gane-McCalla, Dare nate between theories are obtained from binary Institute . Can schedules’ perceived value be choice experiments using Adaptive Design explained by the perceived value of just a few Optimization (ADO), in which the gamble pairs reinforcers? The additive noise model states that are not randomly drawn but rather carefully discounted reinforcer value simply adds together selected to maximally discriminate the theories linearly but as time passes noise is added, V under consideration. Data from a simulation O = Overall value = P vi where vi = effective or study will be presented in which the fanning perceived value of a reinforcer at time i Our uni- out hypothesis is tested by comparing Expected fied theory integrates initial value of outcomes, Utility and Weighted Expected Utility. delay and risk. Results from samples suffice to Tuesday, 10:00 (Session C) characterize entire schedules. Three derivatives Does Hierarchical Complexity of of immediate reinforcer value with respect to Items Predict Synchrony Across time of a reinforcer summate many properties of Content and Gaps Between Stages? discounting accounts of reinforcement schedules. , Harvard Medical School, A trial consisted of a two chain schedule. The

21 first link consisted of the presentation of one where D for two factor values is defined as of a large number of samples from a t schedule (E(|A − B|p))1/p, with A and B being the ran- (Schoenfeld & Cole, 1972). The second link was dom variables (or some measurable transforma- a choice between a left key indicting the sample tions thereof) corresponding to these factor val- was lean or the right key indicating it rich. ues. If the inequality is violated for at least one µ ν µ ν The value of an instantaneous reinforcer is Ai. choice of (x ,u ), (y ,v ), p, and of the measur- The perceived sample value was an hyperbolic able transformations, selective influence is ruled function of how soon before choice a single out. reinforcer was (first derivative) as the Commons et al. (1982)/Mazur’s (1987) equation for delay, Monday, 9:00 (Session C) v (delay) = A /(1 + k d ). Risk is the derivative Fundamental Properties of Reverse Haz- i i 1 i Richard Chechile of Commons et al. (1982)/Mazur’s (1987) ard Functions. , Tufts equation for delay: v (risk) = −A k /(k d +1)2 University . The hazard function h(x) for a con- i i 2 2 i − was well fit by a negative power function as tinuous random variable x is defined as f(x)/[1 proposed. F (x)]. There are many applications of haz- ard function analysis in , Sunday, 11:10 (Session C) thereby attesting to the utility of this proba- A Criterion and Tests for Selective Proba- bilistic construction. Yet in stochastic model- bilistic Causality. Ehtibar Dzhafarov, ing there also is a reverse hazard function r(x), ¨ Purdue U , Janne Kujala, JyvSession which is defined as f(x)/F (x). Although less is ¨ AskylSession A U . A general definition and a known about reverse hazard functions and fewer criterion (a necessary and sufficient condition) applications can be found for these functions in are formulated for an arbitrary set of external comparison to hazard functions, there are many factors to selectively influence a corresponding interesting and useful properties of reverse haz- set of random variables, jointly distributed at ard functions that have the potential for appli- every treatment (i.e., a set of factor values cation in psychological research. In this paper, containing precisely one value of each factor). some new results concerning reverse hazard func- The variables are selectively influenced by the tions are developed. corresponding factors if and only if there is a jointly distributed set of random variables, one Tuesday, 16:00 (Session C) variable for every value of every factor, such Changing Minds: Estimation of Multiple that every subset of this set that corresponds Preference States Via Normalized Max- to a treatment is distributed as the original imum Likelihood. Clintin Davis-Stober, variables at this treatment. We call this the joint University of Missouri . Given multiple presenta- distribution criterion for selective influence. It tions of two choice alternatives, decision makers is applicable to any set of random entities (not do not deterministically choose one alternative necessarily numeric) and any set of factors. The over another. I present a “branch and bound” distance tests (necessary conditions) for selective algorithm that: 1) classifies whether an individ- influence previously formulated for two random ual decision maker is making choices according variables in a two-by-two factorial design are to a single or multiple preference state(s) consis- extended to arbitrary sets of factors and random tent with a pre-specified theory, and 2) estimates variables. A distance test consists in choosing a an optimal mixture distribution over those pref- number p ≥ 1, two distinct factors, µ and ν, and erence states using normalized maximum likeli- two distinct values (xµ,uµ)and(yν,vν) for each hood. I re-examine several existing datasets col- of them, and ascertaining whether lected to test a variety of axioms and paradoxes using this methodology. max {Dxµyν ,Dxµvν,Duµyν,Duµvν}≤ 1 Sunday, 12:10 (Session B) (Dxµyν + Dxµvν + Duµyν + Duµvν ) , 2 The effect of losses on maximization:

22 Loss aversion versus a global-effect model. tainty across beliefs is inherently represented and Eldad Yechiam, Technion - Israel Institute thus the expected uncertainty reductions for can- of Technology . Previous findings have shown didate stimuli can be evaluated. Bayesian active that losses increase performance level in abstract learning models offer the prediction that an ac- tasks compared to rewards. Additionally, it has tive learner would select the stimuli for which been shown that losses increase maximization: the expected uncertainty across all hypotheses is The so called “successful loser” effect (Bereby- minimized. This research contrasts four possible Meyer & Erev, 1998). The current paper exam- hypothesis spaces for active learning consisting ines three simple quantitative models that ac- of two simple cue-combination models and two count for the effect of losses on maximization. possible priors. To tease apart the models, an au- The first model involves loss aversion whereby tomated search of associative learning structures there is an asymmetry in the weight of losses for which the models make maximally different compared to gains (following prospect theory). predictions was performed. Human participants The second and third models are global-effect were tested on these same learning structures in models assuming that losses increase the sub- the context of an allergy diagnosis task. At var- jective weight of all outcomes associated with ious points in training, participants were asked the loss-producing alternative, including gain which cues they would find the most informative components. An additive global-effect model to learn about; i.e., their active learning prefer- (whereby payoff is increased by a positive pa- ences were assessed. The model and prior com- rameter c) is compared to a multiplicative global- binations that best mimic human active learning effect model (whereby payoff is multiplied by a will be discussed. parameter c with a value greater than 1). The predictions of the three models regarding the nec- Monday, 11:10 (Session B) essary conditions for the successful-loser effect Computer-Based Assessment of Planning are examined in three published datasets (Ert & and . Holly Jimison, Stu- Yechiam, 2007; Haruvy & Erev, 2002, and Levin, art Hagler, Misha Pavel, Oregon Health & Weller, Pederson, & Harshman, 2007). The re- Science University . Planning abilities are a key sults show that the only model consistent with component of cognitive health and important to the pattern observed in all three studies is the problem solving. These abilities are often com- multiplicative global-effect model. The adaptive promised with brain injury, stroke, and neurolog- advantage of the global-effect model is that it ical disease. Computerized, model based assess- does not assume that absolute losses impair ex- ment of these abilities can be used in care and ploration or maximization (compared to relative rehabilitation. To address this issue, we devel- losses) when they are presented along with equal oped a research version of the popular computer or larger gains. game FreeCell that enabled us to manipulate the planning difficulty. This computer card game is Monday, 9:20 (Session B) a version of Solitaire where the initial state is Exploring Active Learning in a Bayesian a random arrangement of cards in 8 work piles. Framework. Stephen Denton, John Kr- The player must move these cards ‘constrained uschke, Indiana University . Bayesian ap- by the order and by suit’ and employing 4 free proaches to learning involve incrementally updat- piles in such a way that the resulting state is 4 ing degrees of belief across a space of hypotheses home piles of sorted cards. We adapt the ini- whenever an observer passively observes a stimu- tial card layout to the skill level of each indi- lus and outcome. But these approaches also pro- vidual user and monitor the user’s efficiency in vide a framework for models of active learning moving toward solution by comparing our auto- – learning in which stimuli are actively probed mated solver algorithm to human performance. to disambiguate potential beliefs regarding out- The planning ability is then estimated to be the comes. Within a Bayesian framework, uncer- efficiency of the player relative to the ideal so-

23 lution compared to random guessing. By mon- tive interactions. We believe this research pro- itoring the naturalistic game play in the home gram can provide a new unifying strand between over time in this way, we are able to character- mathematical biology—evolutionary game the- ize within-subject trends in planning or executive ory and population dynamics—and cognitive sci- function, as well as variability in the cognitive ence. measures over time. To summarize the results Tuesday, 9:40 (Session B) we developed a metric that appears to be corre- Modeling Multitrial Free Recall. lated with traditional tests of cognitive abilities. James Pooley, Michael Lee, UCI , William We have demonstrated our ability to keep users Shankle, Medical Care Corporation .We engaged with this game and have integrated it develop a simple psychological measurement into a cognitive health coaching intervention. model, based on the SIMPLE model (Brown, Neath, & Chater 2007), for multitrial free recall Monday, 15:40 (Session C) tasks in which individuals are presented with Evolution and cognitive cost of ethnocen- identical words with word order fixed across the trism. Artem Kaznatcheev, McGill Univer- tasks. Existing models for multitrial free recall sity . There is mounting evidence that ethno- have typically assumed randomized intertask centrism - commonly thought to rely on com- word order, and we compare our model with plex - may arise through biolog- these in terms of its ability to account for ical evolution in populations with minimal cog- various serial position curves. We discuss the nitive abilities. We use the tools of evolutionary importance of fixed versus randomized intertask and computational modeling to ex- word order in multitrial free recall tasks, and amine the evolution of ethnocentrism in a com- discuss the potential use of our model in a petitive world. We create a competitive environ- clinical setting. ment through Prisoners dilemma interactions be- tween agents in a spatial lattice where groups are Monday, 16:20 (Session A) marked by arbitrary tags uncorrelated to strat- A Bayesian Analysis of the Wisdom of egy or location. Ethnocentric agents are able to Crowds Within. Brent Miller, Michael differentiate between in- and out-group partners, Lee, Mark Steyvers, University of Califor- and adjust their behavior accordingly (cooperat- nia, Irvine . When averaging estimates from in- ing with in-group and defecting from out-group). dividuals, the aggregate often comes closer to the Humanitarian or selfish agents never make this true answer than any individuals guess. Assum- distinction and always cooperate or defect, re- ing an individuals guess is drawn from amongst spectively. Thus, ethnocentric agents can be seen a range of potential responses, would an aggre- as more cognitively complex, and we associate a gation of several of these responses also exhibit fitness cost with this complexity. We test the ro- a “wisdom of crowds” effect? Vul and Pashler bustness of ethnocentrism, concluding that eth- (2008) found evidence for this using tasks where nocentrism is not robust against increases in cost individuals gave simple numerical estimates. In of cognition. This suggests that the underlying this study, we examine this wisdom of crowds mechanism behind ethnocentrism cannot be as effect for a series of ranking tasks, where the complex as previously believed. Further, we ad- goal is to reconstruct, from memory, the order of dress several hypotheses on the evolution of eth- time-based events or the magnitude of physical nocentrism in spatial models. We confirm that properties. Each task is performed twice, with humanitarians are suppressed largely by ethno- an intervening distractor task. We combine the centrics. Paradoxically, we observe that the pro- two estimates from each individual to form a wis- portion of cooperation is higher in worlds domi- dom of crowds aggregation. We then introduce nated by ethnocentrics. Suppressing free-riders, a Bayesian measurement model to describe indi- such as selfish and traitorous agents, allows eth- viduals performances across questions and across nocentrics to maintain higher levels of coopera- repetitions, and to measure differences between

24 the individual responses and the aggregated esti- a model of simple response time (RT) in which mates. In results consistent with Vul and Pash- activation is accumulated linearly through a cas- ler, we show that an individuals second estimate cade of units, with a response being made when generally does not improve upon the first, but activation in the terminal unit reaches a crite- that the aggregate of the two is more accurate rion. They showed, assuming IID trial-to-trial than either individual response. variation in the rates and the criterion, that the limiting RT distribution predicted by this Monday, 11:30 (Session B) type of evidence accumulation model is Lognor- Toward a Cognitive Model of a Computer- mal. I propose an extension, the Lognormal Based Trail Making Test. Kyle Ambert, Race (LNR) model, which accounts for choice Holly Jimison, Misha Pavel, Oregon Health accuracy and RT distribution as the outcome & Science University . The Trail Making Test of a race between cascades of linear accumu- (TMT) is a two-part neuropsychological assess- lators corresponding to each possible response. ment used as a measure of visual search and The tractability of this model enables derivation motor-control ability (TMT-A), as well as an in- of computationally tractable expressions for the dex of cognitive set-switching capacity (TMT-B). likelihood of correct and error responses for any Although ubiquitous in the clinical practice and number of potential choices in the case where neuropsychological research, few extensions to trial-to-trial variations in rates are correlated the TMT exist with the ability to track changes across accumulators. I examine the ability of the in player’s performance over time. In this pa- LNR model to account for benchmark choice RT per we describe a version of the TMT that was phenomena with a focus on the role of rate cor- adapted as a computer game, allowing essentially relations and their psychological causes. an unlimited number of presentations and sup- porting variable degrees of difficulty. This com- Monday, 11:50 (Session B) puter game, called Scavenger Hunt, was included Generalized Assessment of Divided At- in a longitudinal study that allowed us to as- tention Through Unobtrusive Monitoring sess and monitor the performance of elder par- of Computerized Tasks. James McKanna, ticipants over more than one year. The data Holly Jimison, Misha Pavel, Oregon Health were analyzed using a simple additive model of & Science University . This paper describes a the contribution of the different cognitive com- computational modeling effort that is a part of ponents required to accomplish the game tasks. a larger study of elders cognitive functions over For example, we hypothesized that the time sub- time. Here we specifically highlight divided at- jects took to complete the first move in a game tention, the ability to allocate focus to multiple included time spent encoding that game’s search tasks at once; this is a key cognitive function string into memory. Thus, we estimated the which decreases with age and which determines contribution of short-term memory to Scavenger performance on a variety of everyday activities Hunt game performance by comparing the dis- like driving. Although there are numerous exper- tribution of first move times to that of subse- imental approaches to the assessment of divided quent moves. Similar techniques were used to attention, few tools exist to measure it in real estimate the contribution of visual search, mo- life. Accurate measurement of variability in this tor speed, memory retrieval, and cognitive load skill over time, relative to a patient-specific base- due to set switching, ultimately giving us useful line, would allow inference of other forms of cog- insights into the cognitive capacity of each user. nitive decline, most importantly those leading to Mild Cognitive Impairment or Alzheimer’s Dis- Monday, 10:00 (Session C) ease. In this study, we present a generalizable The Lognormal Race Model of Choice RT. method for determining relative divided atten- Andrew Heathcote, University of Newcastle, tion ability through analysis of any repeatable, Australia . Ulrich and Miller (1993) considered multi-dimensional task. We tested this method

25 using an adaptive computer game which requires Theories of polarization that attribute bias to attention in each of two dimensions simultane- the reader do not account for non-monotonicity, ously, played by elders at risk for cognitive de- whereas it follows naturally from a model that cline. Each individual players performance was attributes the bias to the authors. evaluated using two extreme models: a random Tuesday, 11:10 (Session B) player and an ideal player. Performance on each Developing Knowledge and Probing Its of the dimensions was measured during situations Structure with Primed Lexical Decision. of low cognitive load to determine relative atten- Greg Cox, Richard Shiffrin, Indiana Uni- tion paid to each aspect of the game; this ratio versity . Both episodic memory and knowledge was then applied to high-load situations to deter- retrieval are affected by the structure of knowl- mine total attentional capacity. We will present edge at the time of testing. We bring the struc- data from 30 subjects monitored over the course ture of knowledge under experimental control: of one year to show the stability of the results Building upon the thesis research of Angela Nel- over time. son (2009), we train participants over several weeks to match Chinese characters to themselves, Monday, 9:00 (Session B) trying to detect slight physical changes (par- A Rational Model of Attitude Polariza- ticipants have no prior knowledge of Chinese). tion. Adam Darlow, Brown University . The characters are presented in pairs, providing Lord, Ross and Lepper (1979) first demonstrated the opportunity for associations to form inciden- attitude polarization when they showed that tally. Unbeknownst to participants, the pairs are both students supporting capital punishment drawn from a set of associative structures, in- and those opposing it became more extreme in cluding linear chains of varying length (e.g., A-B- their views after reading the same mixed evi- C-D-E) and binary trees. We use primed pseudo- dence. Many explanations of polarization have lexical decision to test whether training has pro- been proposed, all of which attribute cognitive or duced direct (e.g. A-B) or indirect (e.g. A-C) motivational biases to the reader. In this paper, associations. The task presents a non-diagnostic a generative Bayesian model of biased evidence prime character followed by a test character re- demonstrates that unbiased rational readers can quiring a decision whether the test character had still show polarization if they consider the pos- ever been studied. The relation of the prime to sibility that the authors are biased. The model the test character allows the presence of associa- assumes that that evidence is sampled from the tive structures in knowledge to be inferred. We world, but that authors may use biased sam- observe effects of structural distance and num- pling procedures that skew the proportion of ev- ber of related associates. Implications for mod- idence that supports a claim or opposes it. A els of knowledge formation and retrieval are dis- reader making inference in this rational model is cussed. Other transfer tasks involved perception more likely to attribute bias to an author if they and episodic memory: The results, implications, present evidence that is incompatible with their and relation of these tasks to the lexical decision prior beliefs than if they present evidence that findings will be discussed as time permits. is compatible. According to the model, mixed evidence can cause polarization if the mixed evi- Monday, 15:40 (Session B) dence is in the form of different authors present- How to analyze multifactorial neuroimag- ing evidence for different sides of a debate, but ing experimental designs with pattern not if evidence for both sides is presented by a classifiers. Joseph Dunlop, Herve´ Abdi, single author. In addition to polarization, the The University of Texas at Dallas . Multivariate model also demonstrates non-monotonicity. As Pattern Classification Analysis (MVPA) of func- the evidence that an author presents becomes tional neuroimages has recently advanced rapidly more one-sided, it at some point begins to in- to include new types of classification techniques, fluence the readers opinion less instead of more. such as pattern classifiers used on single factor

26 designs. Here, we present a new approach which beliefs, but in the case of judging performance extends these techniques to multifactorial design competitions, one hopes the judges are all fMRI experiments. In functional using the same set of qualitative schemata for imaging, pattern classifiers identify patterns of judging. This qualifies the problem as one that brain activation associated with the levels of an Cultural Consensus Theory (CCT) is designed experimental factor (e.g., assign scans to some to handle. CCT is a method of aggregating the predefined object categories). We evaluate the responses of informants who share a common performance of the classifier for a given subject set of experiences or beliefs. It is in wide use by assigning new scans to one category. From as a tool to determine ’consensus’ from small the classifiers pattern of errors, we compute a groups in cultural anthropology (e.g. Batchelder category by category distance matrix for each & Romney, Psychometrika, 1988; Romney, subject, where the entries are d0 values (because Weller, & Batchelder, American Anthropologist, d0 is obtained from error rates, the technique 1986). One aspect of CCT is that it provides works with any pattern classifier). We analyze a statistical measure of whether or not the the distance matrices using DISTATIS, which is judges’ responses appear to follow from a shared a generalization of MDS. Like MDS, DISTATIS unitary perspective indicating a consensus. This converts the distance matrices to semi-positive paper compares CCT and a number of the definite matrices, which are then integrated into ranking algorithms on a large body of data from an optimal weighted average called the “compro- ballroom dance competition, both when there mise.” We eigendecompose this compromise into appears to be consensus and when there is not. a set of factor scores (representing the coordi- nates of the categories) and project the subjects Sunday, 11:10 (Session B) matrices onto these factors. In the same fac- How many tools to include into an adap- tor space, we project linear contrasts represent- tive toolbox? A Bayesian approach. ing specific hypotheses, main effects, and interac- Benjamin Scheibehenne, Jorg¨ Rieskamp, tions. Finally, the reliability and quality of these University of Basel . A common assumption effects are evaluated using cross-validation and within many research areas in psychology is that resampling techniques, and expressed as multi- people possess a repertoire of strategies to solve variate confidence, tolerance, and prediction el- the problems they face. This strategy repertoire lipsoids. approach, sometimes referred to as an “adaptive toolbox”, provides a fruitful way to explain intra- Monday, 15:40 (Session A) and inter-individual differences in cognitive pro- Rank-Aggregation and Consensus in cesses. However, this approach suffers from the Ballroom Competition. Royce Anders, problem that the number of assumed strategies William Batchelder, UC Irvine . In a large is often not constrained a-priori. Thus, a re- number of situations, items (e.g. competitors, searcher may extend a given toolbox with ad- candidates for election, courses of action) are ditional strategies post-hoc to improve the fit to ranked by two or more judges, and an aggre- the data. This makes it difficult to rigorously test gated ranking is desired. There are various and compare this class of models. To prevent this algorithmic ways to achieve an aggregated rank- problem, a criterion is necessary to decide how ing (Borda, Kemeny-Young, Schulze, Skating many strategies a toolbox should include to de- System methods), however unsurprisingly they scribe a given set of choice data. Here, Bayesian do not always produce the same aggregate. statistics provide a useful path to evaluate tool- This topic is investigated in ballroom dance boxes of different size based on their marginal competitions, characterized by a small number likelihood. The present work illustrates how such of items (around 6) and judges (5 to 15). techniques can be fruitfully applied within a de- Rank aggregation algorithms are often created cision making context. The method we propose to handle qualitative differences of experts’ is implemented in BUGS language and utilizes

27 and compares hierarchical models that mimic dif- one or two perspective images. Furthermore, due ferent cognitive toolboxes. We demonstrate the to the fact that objects are opaque, the observer applicability of our approach by means of param- can “see” only the front surfaces of the objects eter recovery studies and based on actual empir- (Marr’s 2.5D sketch). So how is the observer able ical data taken from previously published exper- to form a complete map of a room, including the iments. The present work also makes the novel back, invisible parts of the objects, the objects contribution of showing how Bayesian techniques that are partially occluded, and the spaces be- can be used to test adaptive toolboxes of various hind the objects? We have already shown that, sizes and complexity against alternative decision contrary to Marrs claim, the entire 3D shape, in- making theories. cluding its front as well as back parts, can be recovered from a single 2D retinal image. The Monday, 9:40 (Session B) human visual system does it by applying a priori A multivariate hierarchical Bayesian constraints (3D symmetry, compactness and pla- model of confidence and accuracy in prob- narity of contours). Once the entire 3D shapes in abilistic category learning. Ed Merkle, the scene are recovered, it is possible to recover Wichita State University, Matt Jones, Uni- the spaces between them and form a complete versity of Colorado, Winston Sieck, Applied spatial map. In the experiment, subjects viewed Research Associates . In this talk, we describe a room, containing several pieces of furniture, a multivariate model that describes the time from a single viewing position, and were asked course of both accuracy and confidence in to draw a top view of the room. The ground repeated choice tasks. The modeling framework truth was established by the PhaseSpace system, we develop has interpretable parameters, ac- whose maximal absolute error in reconstructing counts for individual differences, simultaneously positions of 3D points was 2cm. The subjects describes growth in confidence and accuracy, performance was very accurate and similar to describes relationships between confidence and that of a computational model. accuracy, and describes differences between ex- perimental conditions. We illustrate the model Sunday, 15:40 (Session C) using data from a probabilistic category learning Cognitive Modeling as Derived Measure- experiment, in which subjects were asked to ment. Chung-Ping Cheng, National Chung distinguish between two artificial diseases based Cheng University, Ching-Fan Sheu, National on three binary symptoms. Five experimental Chung Cheng University . Parameters in the conditions were implemented that differed in the quantitative model of a cognitive process are learning strategies subjects were directed to use. often interpreted to reflect individual cogni- The model isolates specific effects of the learning tive differences and are compared across groups. strategies on starting points, growth rates, and Researchers often do not distinguish variables, asymptotes of confidence and accuracy. which are directly observed by apparatus or ques- Sunday, 16:00 (Session B) tionnaire, and parameters, which are indirectly Human recovery of the shape of a 3D obtained. In the study, a definition of scale type scene. Taekyu Kwon, Yun Shi, Tadamasa for model parameters is proposed by treating Sawada, Yunfeng Li, Joseph Catrambone, cognitive modeling as a derived measurement. Stephen Sebastian, Zygmunt Pizlo, Purdue As a consequence, it becomes possible to deter- University . Our everyday life experience sug- mine which statistical procedures are appropriate gests that we are able to form a complete and for comparing model parameters across groups. accurate spatial map of the environment (e.g. In light of the derived measurement approach, an office or a living room) from a single view- it is shown that parameters of cognitive mod- ing position. This problem is computationally els, occasionally, may not be quantitative. With quite difficult because the observer has to recon- illustrative examples from cognitive psychology, struct 3D metric structure of the scene from only we demonstrate how to assign scale types to pa-

28 rameters of cognitive quantitative models. We Webb, James Webb, UCL . In a lexical-decision also provided justification of the proposal to as- task (LDT), subjects make speeded responses to sign scale type to parameter based on meaning- whether a letter-string is a word or non-word. fulness. LDT is extensively employed throughout experi- mental psychology, particularly in psycholinguis- Sunday, 15:20 (Session B) tics. In the typical analysis of LDT data in these Veridical binocular perception of met- contexts, however, accuracy and response latency ric properties of 3D shapes. Yun Shi, are analyzed independently, often with aggrega- Taekyu Kwon, Tadamasa Sawada, Yun- tion across subjects and/or items. In this work, feng Li, Zygmunt Pizlo, Purdue University we present a hierarchical model that accommo- . It has been shown that binocular perception dates both the coupled nature of latency and of depth intervals is both inaccurate and unreli- accuracy, as well as subject-level and item-level able, but that of depth order (called stereoacuity) variability in these outcomes. Omittingfull de- is extremely reliable. Our recent psychophysi- tails, if y and z are respectively the latency cal experiments showed, for the first time, that ji ji and response to the ith word stimulus presented human binocular 3D shape recovery of symmet- to subject j, we model these as ric polyhedra is extremely reliable and accurate. This contrasts with poor performance of exist-  W(a0 + b0ψji,λ0,θ0) , if zji =0, ing models of binocular shape perception, in y ∼ ji W(a + b ψ ,λ ,θ ) , if z =1, which the 3D shape is reconstructed from the 1 1 ji 1 1 ji ∼ information about depth intervals. These re- zji Bernoulli (pji) , −1 sults strongly suggest that human binocular per- pji = logit (az + bzψji) , ception of shape is not based on reconstructing depth intervals. Our new model combines the where W (µ, λ, θ) denotes a Wald distribution 3D symmetry prior with the sensory information with mean µ, shape λ and offset θ. Here, ψji is a about depth order of the object points. Specif- latent-variable that is a subject-varying function T ically, a single 2D orthographic image of a 3D of lexical items, i.e. ψji = αj + βj xji, where xji symmetric object determines the 3D shape of denotes the values of relevant lexical-variables of this object up to only one unknown parameter, stimulus ji, and αj and βj are drawn from drawn the aspect ratio of the shape. 3D shapes with from common subject-wide distributions. Using different aspect ratios are likely to have differ- a large LDT data-set, we compare this model ent depth order of their vertices. The model as- with the more standard methods of analyzing sumes that the true depth order is known from the roles of variables (frequency, length, ortho- stereoacuity, and this ordinal information is used graphic regularity, etc) on lexical processing. We to restrict the range of aspect ratios of the sym- demonstrate how the coupled-outcome hierarchi- metric 3D shape. To verify the role of ordinal cal model can considerably clarify, while the non- information about depth in recovering metric 3D hierarchical independent analysis can often dis- shape, we compared the subjects performance in tort, the roles of these lexical variables. stereoacuity test with that in 3D shape recov- Monday, 15:20 (Session B) ery task. Subjects with poor stereoacuity (high Distance-based Partial Least Squares Re- threshold) are less reliable in 3D shape recovery gression (DISPLS): A new approach to task. We conclude by discussing an elaboration relate distances in psychology and neu- of our binocular model so that it does not rely roimaging. Anjali Krishnan, The Univer- on any depth information at all. sity of Texas at Dallas, Nikolaus Kriegesko- Tuesday, 12:10 (Session A) rte, University of Cambridge, Herve´ Abdi, The A Hierarchical Logistic-Wald Model for University of Texas at Dallas . How do we per- the Analysis of Lexical-Decision Data. ceive the world? One possible explanation is Mark Andrews, NTU and UCL, James that the brain associates semantic (high-level)

29 information with perceptual (low-level) informa- optimally combining the available information tion. Researchers have tried to detect this se- from episodic memory and prior knowledge. In mantic information through various behavioral, this work, we extend these rational models of neuroimaging, and computational approaches. A memory by incorporating individual differences. common feature of these approaches is the no- We assume that individuals can have varying tion of distance, which quantifies similarity (or internal noise levels and also different subjective dissimilarity) between stimuli. We could use dis- priors. We estimate these individual differences tances between stimuli based on their perceptual using a Bayesian data analysis applied to the features to predict distances between the same rational model. Therefore, in this approach, stimuli based on behavioral or neural measures. we combine two different kinds of Bayesian The prediction could be interpreted as the low- analysis. From the perspective of the individual level information encoded in the brain, which we performing the task, we use a Bayesian analysis obtain from the stimuli. The residual of the pre- to motivate a rational model of memory. From diction could be interpreted as the high-level in- the perspective of the experimenter, we use the formation encoded in the brain, which we do not outcome of this Bayesian inference as a forward obtain from the stimuli. Our research question (generative) step to analyze the observed data involves prediction, but standard methods used in the experiment. We illustrate this approach to analyze distances such as multidimensional with an empirical study in which people recall scaling and procrustes analysis are not predic- the height of females, males and ambiguous tive. Partial Least Squares Regression (PLSR) silhouettes. is a method that can predict a set of depen- dent variables from a set of independent vari- Monday, 12:10 (Session C) ables, but PLSR does not analyze distances. We Learning of dimensional biases in human propose a new statistical method called Distance- categorization. Adam Sanborn, University based Partial Least Squares Regression (DIS- College London, Katherine Heller, Univer- PLS), which predicts distances from distances. sity of Cambridge, Nick Chater, University We illustrate DISPLS with data provided by College London . Existing models of categoriza- Nikolaus Kriegeskorte, which were collected on tion typically represent to-be-classified items as humans (fMRI) and monkeys (single-cell record- points in a multidimensional space. While from a ings). We will show how DISPLS could model mathematical point of view, an infinite number of what the brain encodes as the sum of the low- basis sets can be used to represent points in this level information predicted by stimuli and the space, the choice of basis set is psychologically high-level information unique to the brain and crucial. People generally choose the same basis also show how this information could be pre- dimensions, and have a strong preference to gen- served across species. eralize along the axes of these dimensions, but not ‘diagonally’. What makes some choices of Monday, 12:10 (Session A) dimension special? We explore the idea that the Estimating Individual Differences dimensions used by people echo the natural vari- in a Rational Model of Memory. ation in the environment. Specifically, we present Pernille Hemmer, Mark Steyvers, UC a rational model that does not assume a partic- Irvine . Previous research has shown that people ular basis, but learns the same type of dimen- can incorporate prior knowledge when recalling sional generalizations that people display. This events from episodic memory. Our earlier bias is shaped by exposing the model to many modeling work has explained effects of prior categories with a structure hypothesized to be knowledge with a rational model of memory. In like those which children encounter. The learning this approach, it is assumed that individuals behavior of the model captures the developmen- solve the computational problem of memory re- tal shift from roughly ‘isotropic’ for children to call (what was the original stimulus shown?) by the axis-aligned generalization that adults show.

30 language corpora from -directed and child- directed speech. These representations were then Tuesday, 10:50 (Session B) compared with a free association network con- Modeling knowledge formation and sisting of the same subset of words. The analy- frequency effects in lexical decision. ses computed graph similarities based on edit- Angela Nelson, UCSD, Richard Shiffrin, distance and compared other statistical prop- Indiana University . In this talk we present erties of the whole networks. Early and late work from two previous studies showing the learned words were also analyzed for their statis- effects of differential training experience on tical properties in the two representations. The novel items in a pseudo-lexical decision task. results revealed statistical properties of the early The first experiment uses a training paradigm in language environment that may be consistent which contextual information plays a large role with an associative structure in child-directed in knowledge development and varies based on speech. item frequency; the second experiment however trains all items in a relatively context-free envi- Sunday, 15:20 (Session C) ronment. Training frequency is shown to play Questionnaire data are linear enough: In- a large role in the pseudo-lexical decision task ferences about interactions based on poly- for both experiments, including (in the second tomous items. Annemarie Zand Scholten, experiment) when subjects are re-tested after a Denny Borsboom, University of Amsterdam . six-week delay: high frequency items are always Interaction effects found on psychological abil- recognized more quickly and more accurately ity tests can be at risk for invalid inference un- than low frequency items. Given that contextual der specific circumstances. Although these cir- information during training only varied based cumstances are not extremely frequent, they do on frequency in experiment 1 an explanation occur regularly in and of these effects based on context alone would cannot always be avoided. Critical factors are not predict the results seen in experiment 2. very high (or low) test difficulty, high item dis- We present a portion of the SARKAE model crimination, long tests, large effect size and small that accounts for the development of knowledge sample size. These factors, however, have shown during the two types of training and the use of to increase the risk of ambiguous results for di- that knowledge in the pseudo-lexical decision chotomous items only. With a simulation study task. The model is able to successfully account we show that when items are polytomous, risk for the results shown in both experiments 1 and of inferential error much smaller. The graded re- 2. sponse model was used to simulate data. If the relationship between an underlying quantitative Tuesday, 11:30 (Session A) variable and observed data can be described by Semantic Associates in Motherese: Com- such a model, questionnaires using three or more paring the Associative Structure of Adult response categories will be relatively safe from and Child-directed Speech with Graph inferential error concerning interactions due to Analysis. Thomas Hills, University of Basel arbitrary choice of measurement scale. . Does child-directed speech reveal transitional probabilities consistent with a more ‘free associa- Sunday, 11:30 (Session B) tive’ speech pattern than adult-directed speech? Judgment Field Theory: A computa- If so, this associative motherese would help ex- tional model of subjective probabilities. plain why associates predict order of acquisition Timothy Pleskac, Michigan State University in early word learning, and provide clues to a . Psychological theories of subjective probability cognitive process that would generate motherese judgments (SPs) typically assume that accumu- and possibly facilitate learning. To test the main lated evidence (support) mediates the relation- question, semantic space representations based ship between the to-be-judged event and the SP on word co-occurrences were computed using (e.g., Tversky & Koehler, 1994). To model the

31 time course of this evidence accumulation I have vague – and occasionally ad hoc – metrics for integrated decision field theory’s (DFT; Roe, et linking learning to looking. But, when data al., 2001) dynamic network of attribute process- is sparse and linking functions are imprecisely ing with Pleskac and Busemeyer’s (in press) dif- specified, conclusions are hotly debated and fusion model of confidence. The model–judgment failures to replicate are common. We propose a field theory (JFT)–treats events (Lance winning model selection framework for choosing explicit a bicycle race) as composed of a set of attributes linking hypotheses in a formal, consistent man- (Lance’s climbing and sprinting ability). Dur- ner. The entire set of eye movements generated ing deliberation, attention shifts between the by a participant performing a learning task is attributes. At each moment, the attended-to- treated as a time-series. Candidate models are attribute results in a level of evidence support- then assigned likelihoods in proportion to their ing each event in the evaluation set relative to probability of generating these time-series. We the remaining events, which is then accumulated. validate this approach by determining its rate of Markers–one for each estimate–are placed across false alarms using a null generating process, and the evidence space. When evidence passes a then apply it to two different infant learning marker there is a probability the judge stops and paradigms—statistical word-learning (Smith & gives the respective estimate. The markers are Yu, 2008) and cued attention (Wu & Kirkham, positioned with the starting point of evidence ac- accepted). In each case, our framework allows cumulation resting at the ignorance prior prob- us to make principled inferences about what ability. JFT offers a single process account of (and how) each infant learns. effects due to a bias towards an ignorance prior; why binary complementarity of SPs occurs; and Tuesday, 11:50 (Session C) why subadditivity of packed hypotheses occurs. Explaining Order Effects. The Belief- JFT also makes new predictions regarding how Adjustment Model Versus The Quantum the events in the evaluation set can interact to Inference Model. Jennifer Trueblood, produce violations of independence. More gen- Jerome Busemeyer, Indiana University . Or- erally, JFT in combination with DFT offers a der of information plays a crucial role in the pro- single process account for judgment and decision cess of updating beliefs across time. Specifically, making rather than a process for judgment and an order effect occurs when probability judg- a process for decision. ments are influenced by the order of a series of in- formation. These order effects are non-Bayesian Sunday, 10:00 (Session B) by nature and are difficult to explain by clas- Linking Learning to Looking: Model sical probability models. We use the empirical Selection for Eye Movements. results of a jury decision-making task conducted Daniel Yurovsky, Indiana University, by McKenzie et al. (2002) to motivate the de- Shohei Hidaka, Japan Advanced Institute of velopment of a quantum inference model for or- Science, Chen Yu, Linda Smith, Indiana der effects. The quantum model is derived from University . Because eye fixations are produced the axiomatic principles of quantum probability rapidly in response to stimuli, gaze patterns can theory and provides a more coherent explana- be a particularly sensitive tool for measuring tion of order effects than heuristic models such as learning. Unfortunately, inferring learning from the belief-adjustment model (Hogarth and Ein- looking in many tasks is anything but straight- horn, 1992). To further test the quantum infer- forward – especially when no other measures ence model, a new jury decision-making experi- provide convergent evidence. This problem is ment is developed. This experiment extends the particularly vexing for investigators studying work of McKenzie et al., and the results are used development, for whom eye movements are often to compare the quantum model to the belief- the only available source of information about adjustment model. Specifically, we compare the learning. As a result, researchers often adopt quantum inference model to two versions of the

32 belief-adjustment model, the adding model and Melissa Prince, Andrew Heathcote, The the averaging model. We show that both the University of Newcastle . Accounts of recognition quantum model and the adding model provide memory data are often based on signal detec- good fits to the data. To distinguish the quantum tion models, where studied (old) and unstudied model from the adding model, we develop a new (new) items are represented as two Gaussian dis- experiment involving extreme evidence. The re- tributions located on a single memory strength sults from this new experiment suggest that the dimension. It is commonly found that the old belief-adjustment model faces limitations when distribution is more variable than the new (i.e., accounting for tasks involving extreme evidence zROC slope=0.8). Ratcliff, Sheu and Gronlund whereas the quantum inference model does not. (1992) suggested that this greater variance might in part be due to variations in the time between Tuesday, 15:40 (Session B) study and test (lag) among old items. We tested A distributed representation of internal this explanation by manipulating variability in time leads to scale invariant timing be- lag by varying (between-subjects) the order in havior. Karthik Shankar, Marc Howard, which study items were tested, such that lag vari- Syracuse University . We present an extension to ability was (a) minimised by testing items in the the temporal context model (TCM) that explic- same (Forward) order as their study presentation itly stores temporal information about when a or (b) maximised by testing in the reverse (Back- stimulus was encountered in the past. The tem- ward) order. Although the lag standard devia- poral information is constructed from a set of tion differed by a factor of more than 4 between temporal context vectors adopted from TCM. We these conditions, maximum likelihood ROC anal- show that an array of temporal context vectors ysis of each participants confidence ratings re- with different time constants is closely related vealed only slightly greater variability for old to the Laplace transform of real time events. items in the Backward compared to the Forward A simple excitatory/inhibitory band of feedfor- condition. These results suggest that, at least for ward connections from the array of temporal the shorter (32 item) study lists used in our ex- context vectors is sufficient to approximately re- periment, lag variation is at best a minor source construct the timing information contained in of old-item variability in memory strength, de- the array of temporal context vectors. The re- spite the old item standard deviation being 1.4 construction procedure has the property that times the new item standard deviation. We will the reconstructed timing of events further in also report results from a follow-up experiment the past is less accurate than the reconstructed with an improved effect size (by increasing list timing of more recent events. We show for- length) and improved power (by using a within- mally that the procedure yields the scalar prop- subjects manipulation of lag variability). erty in the variability of the underlying tim- ing distribution. This mathematical property Tuesday, 16:00 (Session A) leads to wide-ranging implications for a variety Models of Information Integration of paradigms. We show that it leads naturally in Perceptual Decision Making. to explain the scalar property widely-observed in Jared Hotaling, Indiana University, An- timing paradigms. It also accounts for the lack of drew Cohen, University of Massachusetts, a temporal scale in associative learning. Ongoing Jerome Busemeyer, Richard Shiffrin, work suggests that this formalism can be adopted Indiana University . In cognitive science there is to a wide range of phenomena potentially leading a seeming paradox: On the one hand researchers to a unified theory of recency and contiguity in studying judgment and decision making (JDM) episodic memory and timing phenomena. have repeatedly shown that people employ Tuesday, 10:00 (Session B) simple and often sub-optimal strategies when The Contribution of Study-Test Lag to Old integrating information from multiple sources. Item Variability in Recognition Memory. On the other hand another set of researchers has

33 had great success using Bayesian optimal models embedding elements). Within this general archi- to explain information integration in fields such tecture, a formalized syntax for the expression of as categorization, perception, and memory. E(t) discloses consistent family groupings par- One impediment to reconciling this paradox alleling the progression of E(t) values, of which lies in the different experimental methods each the hierarchical positioning of U is a principal group has used. Recently, Hotaling, Cohen, determinant. Marked elevation in E(t) accom- Busemeyer, & Shiffrin (submitted) conducted panies U’s position at a more subordinate level a perceptual decision making study designed of the hierarchy. Such elevation analytically is to bridge this methodological divide and test shown to emanate from nuanced mechanisms of whether the sub-optimal integration found in bin-element access. In addition, uncertainty as verbal problems stated in terms of probabilities ‘information withheld’ is mathematically distin- may also appear in perceptual tasks. Their guished from uncertainty as ‘event unpredictabil- results indicate that a classic JDM finding, the ity’. Potential implications of uncertainty reduc- dilution effect, does arise in perceptual decision tion for clinical stress management are discussed. making. Observers were given strong evidence X favoring A over B, and weak evidence Y also favoring A over B. According to Bayesian Tuesday, 16:00 (Session B) analysis, the odds in favor of A should be Rats, non-rewards, and decaying uncer- multiplied, resulting in an increased likelihood of tainties. Chloe Bracis, James Anderson, A. Instead, Hotaling et al. found that the weak University of Washington, Andrew Goodwin, evidence diluted the strong evidence, producing Environmental Laboratory, ERDC . Phenomena decreased judgments and choice probabilities such as extinction and spontaneous recovery are favoring A, given X & Y, than given X alone. found in both Pavlovian conditioning and instru- I review these empirical findings and test both mental learning. While changes in response rate rational and cognitive models of the integration reflect the animal’s changing expectation of re- process. ward, only Pavlovian conditioning provides clear events to update expectation in response to no Tuesday, 11:10 (Session C) reward. Here we take a mathematical learning Mathematical Expectation of Threat, Un- model (Anderson et al. CogSci 2010), which op- predictability, and Mechanisms of Stochas- erates in event time, that has been used to suc- tic “Faint Threat” in a Model of Decisional cessfully model these phenomena in Pavlovian Control. Matthew Shanahan, Richard conditioning experiments and apply it in an in- Neufeld, The University of Western Ontario . strumental learning context. The model tracks This presentation addresses issues involved with distal and proximal scales of the expected in- calculating mathematical expectation of threat terval to reward as well as their respective er- E(t) in the context of a normative model of rors in order to combine both estimates through decisional control (Shanahan & Neufeld, 2010, their uncertainties. We handle updating expecta- BJMSP). Three choice conditions are possible tions recursively throughout the interval for non- for the decision maker (DM) at each level of a rewards, translating non-events from real time bin-model prototype of a successive-nesting de- into an event time characteristic of Pavlovian cisional hierarchy: decision-making done by the models. Additionally, the model translates be- DM (free choice - C), external selection revealed tween real time and event time to decay uncer- to the DM (no choice - N), or external selection tainties between sessions; this reverts expecta- not revealed to the DM (uncertainty - U). Sce- tion to the long-term mean when there is no in- narios are named by nodal choice condition in formation, giving rise to spontaneous recovery. increasing order of nesting (e.g., CNU denotes a We present examples of fitting the model to data C condition for bin sets, nesting an N condition and demonstrate how the underlying model dy- for bins, in turn nesting a U condition for threat- namics give rise to the observed behavior. This

34 model synthesizes time and event domain re- time. PPO functions according to an underly- wards into a common framework. ing mathematical model, based on an extension of the General Performance Equation (Ander- Sunday, 11:50 (Session A) son & Schunn, 2000), that accounts for stabil- Reward maximization, drift-diffusion ity of knowledge and skill levels by accounting and inter-response times in instrumental for the temporal distribution of training events. conditioning. Patrick Simen, Fuat Balci, It calibrates learning and decay parameters to Princeton University, David Freestone, a learners unique training history, and predicts Brown University . Drift-diffusion models ac- performance at future time-points based on those count for the complete shape of response time parameters. In its current form, PPO gener- distributions in perceptual decision making ates point predictions which estimate specific lev- experiments. In tasks involving rewards as els of performance at specific points in time. feedback, fitted model parameters have been We know, however, that human behavior varies shown to adapt during performance so as to over time, and that model precision generally de- maximize reward rates. We now show that a creases with longer lead times. It would there- diffusion model with drift equal to a simple fore be useful for trainers to have a measure of function of reward rate also captures the shape this uncertainty so that more informed training of inter-response time distributions from mice decisions can be made. We are exploring two and rats in simple conditioning tasks—tasks in methods for computing prediction intervals to which the time between rewards, and not any provide this additional information. The first other perceptual signal, is the critical indepen- method adds variability into the learning and de- dent variable. This model furthermore provides cay parameter estimates to produce a range of a novel account of hyperbolic response rates predicted performance. The second method ex- as a function of reinforcement rates in variable trapolates future residuals from the known resid- interval tasks, and of states of either rapid or uals of the data that the model was calibrated infrequent responding in fixed interval tasks. against. Both methods have positive and nega- Fixed-interval state transitions are triggered by tive characteristics, which will be detailed in the a hierarchically superior interval timer, which is presentation. Although incorporation of predic- itself a drift-diffusion process with parameters tion intervals into a non-stochastic model such tuned to maximize response time precision. as ours is not straightforward, it is an impor- With this tuning, the high-level timer in the tant step in developing PPO into a useful tool model accounts for the law of scalar invariance for learners and trainers. in interval timing. These results suggest that simple processes of noisy accumulation, governed Monday, 15:20 (Session A) by reward-sensitive feedback controllers, play Averaging and Choosing in the critical roles in response generation across a Mind: Should Judgments Be Based wider range of behaviors than has previously on Exemplars, Rules or Both? been hypothesized. Bettina von Helversen, Stefan Her- Monday, 16:40 (Session B) zog, University of Basel . When judging or Calculating Prediction Intervals for Fu- categorizing objects people may rely on different ture Performance. Kelly Addis, Michael kinds of cognitive processes with similarity- Krusmark, Tiffany Jastrzembski, Kevin based and rule-based processes being the most Gluck, Air Force Research Laboratory, Stuart prominent processes under discussion (Ashby, Rodgers, AGS TechNet . The Predictive Per- Alfonse-Reese, Turken, & Waldron, 1998, Psych. formance Optimizer (PPO) is a cognitive tool Review). Models assuming that both types of designed for learners and trainers to track knowl- processes are used, often also assume that one edge and skill change, and to quantitatively pre- process is chosen, depending on the success dict performance at specified future points in with which the process can solve the task (e.g.

35 Ashby et al., 1998). However, machine learning resources a decision maker allocates to a decision and group decision making research suggests task. An observers resources must be divided that when two judgments are available, it is amongst possible response alternatives, thus often better to take the average of the two resulting in a decreasing number of particles per than to choose one of them—particularly when choice as the number of response alternatives errors are somewhat independent and there is increases. This proposal predicts Hick’s Law for uncertainty about which judgment will be more a model endowed with many particles, consistent accurate (Soll & Larrick, 2009; JEP: LMC). In with the notion of Hick’s Law as an account of a simulation study we investigated in five real statistically optimal decision making. With few world environments whether judgment accuracy particles response time and accuracy predictions was higher if (a) an exemplar-based or (b) a depart from Hick’s Law but are consistent with rule-based strategy was chosen or when (c) their our data sets and the notion of humans as predictions were averaged. Our cross validated limited capacity observers. results show that (except for very small samples) Tuesday, 11:50 (Session A) averaging is more accurate than choosing only Modeling the Acquisition of the Men- a rule-based strategy or an exemplar-based tal Lexicon. George Kachergis, Chen Yu, strategy, choosing the strategy which performed Richard Shiffrin, Indiana University . Sev- better in the training sample (based on its BIC) eral studies have found that adults can acquire and as good as perfectly knowing which strategy word-referent pairings after experiencing a series will be more accurate in hindsight. These results of individually-ambiguous trials (i.e., trials con- suggest that averaging may be a successful taining multiple words and referents). To dis- mechanism within one mind. ambiguate pairings – which many learners do with great success – word-referent co-occurrences Monday, 10:50 (Session C) must be integrated across trials. We discuss a va- A particle filter account of multi- riety of factors that have empirically been found alternative decisions. Guy Hawkins, to affect learning performance, such as pair fre- Scott Brown, University of Newcastle, Mark quency, pairs per trial, temporal contiguity, and Steyvers, University of California Irvine, contextual diversity (how many pairs a given pair Eric-Jan Wagenmakers, University of Ams- appears with during training). Using this data, terdam . Most choice response time research has a variety of associative models were constructed, focused on binary decision making. When there implementing diverse attention, memory, and in- are more than two choice alternatives the typical ference mechanisms. We find that some surpris- result is Hick’s Law, that response time and the ing empirical results (e.g., a null frequency effect logarithm of the number of choice alternatives when pairs of different frequency co-occur) can are linearly related (Hick, 1952). We recently be explained by shifting attention and limited reported results from two experiments that memory. Finally, we demonstrate that although did not conform to Hicks Law for decisions learners do assume mutually exclusive pairings in between many choice alternatives, suggesting some situations (to their great advantage), they a limit to Hick’s Law. Rather than using the can also relax this assumption and learn more typical evidence accumulation framework, we complex mappings. propose an alternative model based on capacity limitation. Our particle filter account is capa- Monday, 11:30 (Session C) ble of predicting Hick’s Law, consistent with Discovering Structure by Learning Sparse existing multi-alternative choice models, as well Graphs. Brenden Lake, Joshua Tenen- as the violations of Hick’s Law observed in our baum, Massachusetts Institute of Technology . data sets all with the manipulation of a single Systems of concepts such as colors, animals, parameter: number of particles. The number cities, and artifacts are richly structured, and of particles is interpretable as the quantity of people discover the structure of these domains

36 throughout a lifetime of experience. Discover- features over all contexts it can be in; however, ing structure can be formalized as probabilis- in any particular context, the object has a finite tic inference about the organization of entities, number of features. This is crucial because it and previous work has operationalized learning allows the feature representation of an object to as selection amongst specific candidate hypothe- change appropriate to its context, but the infer- ses such as rings, trees, chains, grids, etc. de- ence for features of an object remains tractable. fined by graph grammars [Kemp & Tenenbaum, The model predicts that the same object should 2008, 116(1), 20-58]. While this model makes be represented as a whole or a conjunction of discrete choices from a limited set, humans ap- parts depending on the distribution of parts in pear to entertain an unlimited range of hypothe- its context (parts co-vary yields a whole; parts ses, many without an obvious grammatical de- independent yields the parts as features). We scription. This structural flexibility is crucial confirm this prediction in a series of experiments for learning about domains that are interestingly using perceptual objects (2-d objects and 3-d structured but lack an obvious structural gram- rendered objects) and concepts (novel animals). mar, such as artifacts and social networks. In this paper, we approach structure discovery as Sunday, 16:00 (Session C) optimization in a continuous space of all possi- A Three-Parameter Item Response Model ble structures, while encouraging structures to of Matching. Matthew Zeigenfuse, be sparsely connected. When reasoning about William Batchelder, Mark Steyvers, animals and cities, the sparse model achieves University of California, Irvine . In a common performance equivalent to more structured ap- experimental paradigm, participants are shown proaches. We also explore a large domain of 1000 two lists of items of equal length and asked to concepts with broad semantic coverage and no associate each element of the second list with simple structure. exactly one element of the first. Often in such experiments, we would like to draw conclusions Monday, 11:10 (Session C) regarding both a participant’s ability in the Feature learning as nonparametric domain from which the list items were drawn as Bayesian inference. Joseph Austerweil, well as the difficulty and discriminability of each Thomas Griffiths, UC Berkeley . A fun- item. A common type of measurement model damental problem faced by cognitive systems for these quantities is Item Response Theory is the formation of basic units to represent (IRT); however, the application of IRT requires observed objects. The computational problem that a participant’s responses on the items be of feature learning has two major facets: (1) any conditionally independent given the participant object can be represented by an infinite number and item parameters, a property known as local of features and (2) the feature representation independence, which is clearly violated for the for an object depends on its context. In both matching experiment. In this talk, we present a perception (e.g., Kanisza, 1975) and categoriza- novel extension to traditional IRT which is able tion (e.g., Wisniewski & Medin, 1994), people to resolve this apparent incompatibility. We flexibly use different feature representations for provide illustrations of model performance on the same object in different contexts. We argue both simulated and actual data. that these two facets of human feature learning are interdependent, complementary and arise Monday, 16:20 (Session B) out of a rational solution to the computational Measuring abnormality in neuropsycho- problem. We present a nonparametric Bayesian logical deficits using generalized linear model for feature learning that finds the best mixed models. Rainer Stollhoff, MPI for feature representation for a set of objects out of in the Sciences . Neuropsycholo- an infinite set of possible features. Any object gists often have to assess whether a patient’s per- can be represented by an infinite number of formance in one or more behavioural tests devi-

37 ates fundamentally from that of a control pop- to a decreasing series of “acceptance” criteria. ulation. A commonly used method is to select One important model that is easy to simulate but for each individual a control population that is has a difficult likelihood function is the Wiener matched in terms of possible contributing factors diffusion process for two-choice decisions. We (age, gender, education, ... ), calculate mean and used an ABC method to fit a Bayesian hierar- standard deviation of the matched controls, and chical version of the diffusion model. We first derive an abnormality score based on an appro- refined the technique by fitting data from a sim- priate test statistic (e.g. z-score, modified t-test). ulation study, and then we fit the model to real Shortcomings of this method are a reduction in data. The data were collected in a perceptual- the number of available control samples due to matching task that varied the discriminability the matching process and - in the case of conti- between same/different pairs as well as instruc- nous contributing factors - a somewhat arbitrary tions that emphasized either speed or accuracy discretization of a continuous variable into inter- of responding. We discuss the ease and accuracy vals (e.g. age is discretized into age-bands). Fur- of the ABC approach, as well as its utility as thermore, while most test statistics used assume compared to other methods for model fitting. a normal distribution experimental test results are often non-normal. Generalized linear mixed Sunday, 9:40 (Session B) models (GLMMs) can be used to alleviate these Adaptive Strategy Choice in Cognitive difficulties and construct continuous abnormality Diagnosis: An Application to Frac- scores. Matching can be achieved by a regression tion Subtraction. Yun Jin Rho, James of test scores on continuous contributing factors, Corter, Matthew Johnson, Teachers Col- and the generalization to exponential family dis- lege, Columbia University . Problem solvers who tributions allows to calculate abnormality scores know several different solution methods or strate- for non-normal test scores. Methods to derive gies for a problem may choose a strategy flex- abnormality scores and diagnostic rules based on ibly for each problem in order to lessen cog- GLMMs will be presented and discussed. Fur- nitive load or to maximize accuracy of perfor- thermore, a demonstration of the methods will be mance under time pressure. This idea is ex- given for the case of diagnosing prosopagnosia, a plored, and consideration is given to how this deficit in facial identification, based on individ- issue of multiple-strategy use will affect cognitive ual performance profiles obtained in several face skill diagnosis. The adaptive strategy choice hy- recognition tests. pothesis is proposed, which states that if test tak- ers know more than one solution strategy, they Monday, 11:50 (Session C) will tend to choose the easiest effective strategy Hierarchical Approximate Bayesian Com- for a particular problem. This idea is applied putation. Brandon Turner, Trisha Van to a well-known problem-solving domain—that Zandt, Mario Peruggia, Ohio State Univer- of mixed-fraction subtraction. In this domain, sity . A natural modeling framework for the adaptive strategy choice leads to a hybrid strat- study of individual and group differences is pro- egy (Method C) in addition to the previously vided by Bayesian hierarchical models. However, recognized Methods A and B (Method A: first some models have likelihoods that have proved convert mixed numbers to improper fractions; to be computationally intractable or difficult to Method B: first separate mixed numbers into in- fully implement. This has motivated the use of teger and fraction; Tatsuoka, 1990). The idea Approximate Bayesian Computation (ABC). In of Method C is that people can select strategies ABC, the computation of the likelihood function on each item, using whichever strategy is easier is replaced by a simulation of the model. Af- between Method A and B. Previous cognitive di- ter simulation of the model, a distance metric is agnosis for mixed fraction subtraction has paid computed between the simulated data and the more attention on Method B because of better observed data. This distance is then compared model fit than Method A. However, the present

38 results from multiple regression and mixed effects strate their use on data from a simple cognitive logistic regression analyses show that Method C task. works better than Method A or B to predict the item difficulties. In addition, Method C shows Tuesday, 15:40 (Session C) the best model fit among the three methods un- A Bayesian model of judgment revisions in Sven Stringer der the NIDA (Noisy inputs, deterministic and a social setting. , University Mark Steyvers gate) model. of Amsterdam, , University Of California, Irvine . Estimating quantities can be Monday, 16:00 (Session A) a difficult task. Especially if other people have The Wisdom of Crowds in Solving Bandit different opinions. We studied how anonymous Problems. Shunan Zhang, Michael Lee, others influence initial estimates of participants. UCI . The “wisdom of the crowds” refers to In this talk we present a Bayesian model of such the idea that the aggregated performance of a a revision process. We will also look at individ- group of people on a challenging task may be ual factors that might play a role in the revision superior to the performance of any of the indi- process. viduals. For some tasks, like estimating a single quantity, it is straightforward to aggregate indi- Sunday, 9:40 (Session C) vidual behavior. For more complicated multidi- Concensus among Concensus Methods. mensional or sequential tasks, however, it is not Anna Popova, Michel Regenwetter, so straightforward. Cognitive models of behav- Sergey Popov, UIUC . in ior are needed, to infer what people know from Economics and Political Science has highlighted how they behave, and allow aggregation to be that competing notions of rational social choice done on the inferred knowledge. We provide a are irreconcilable, because of impossibility the- case study of this role for cognitive modeling in orems and computer simulations. Do concensus the wisdom of crowds, using a multidimensional methods disagree so much in reality? We report sequential optimization problem, known as the on a behavioral social choice comparison of bandit problem, for which there are large differ- Condorcet, Borda, Plurality, Negative Plurality, ences in individual ability. We show that, us- STV, Coombs, and Plurality Runoff one, using ing some established cognitive models of peo- several data sets from American Psychological ple’s decision-making on these problems, aggre- Association presidential elections. The empirical gate performance approaches optimality, and ex- findings contradict theoretical expectations. ceeds the performance of the vast majority of in- Behavioral research in social choice may reveal dividuals. many future surprises. Explaining the empirical agreement among social choice outcomes is an Monday, 9:40 (Session C) open problem (Regenwetter, 2009, Perspectives The Statistical Properties of the Sur- on Psychological Science). vivor Interaction Contrast. Joseph Houpt, James Townsend, Indiana University . The Tuesday, 10:50 (Session C) Survivor Interaction Contrast (SIC) is a powerful Structure from Sequence: A Nonparamet- tool for assessing the architecture and stopping ric Bayesian Analysis of Human Sequence rule of a model of mental processes (Townsend Learning. Jerad Fields, Todd Gureckis, and Nozawa, 1995). Despite its demonstrated New York University . Discovering the latent utility, the methodology has lacked a method for structure underlying sequential information is a statistical testing until now. We will briefly de- problem that is central to many aspects of hu- scribe the SIC then develop some basic statis- man cognition (e.g., language, perception). In tical properties of the measure. These develop- this talk, we describe a new account of hu- ments lead to a statistical test for rejecting cer- man sequence learning based on the principle tain classes of models based on the SIC. We ver- of Bayesian model comparison. The model uses ify these tests using simulated data, then demon- nonparametric Bayesian inference techniques to

39 infer a distribution over possible hidden Markov if an x matches both y and y0, then y and y0 model (HMM) structures that are consistent are equivalent (in the sense of always matching with a particular training sequence. We success- or not matching any stimulus together). This fully apply the model to a number of well known property has nontrivial consequences for sev- findings in artificial grammar learning (AGL) eral issues, ranging from the ancient “sorites” tasks (e.g., Gomez, 2002). Finally, we describe paradox to modeling of discrimination by means a novel AGL experiment which explicitly differ- of Thurstonian-type models. We have tested entiates our account from more traditional ap- the regular well-matchedness hypothesis for loca- proaches. In contrast to accounts based on as- tions of two dots within two side-by-side circles, sociative learning principles (e.g., Simple Recur- and for two side-by-side “flower-like” shapes, rent Networks), we argue that human sequence each described by R+Ka cos aθ+Kb cos bθ in po- learning is better explained as a structured in- lar coordinates. In the location experiment the ference process that reflects the uncertainty over coordinates of the dot in one circle were adjusted candidate structures. to match the location of the dot in another circle. In the shape experiment the amplitudes K ,K Tuesday, 16:20 (Session B) a b of one shape were adjusted to match the other Regularities in Dream Social Networks. shape. The adjustments on the left and on the Richard Schweickert, Purdue University, right alternated in long series according to the Charles Viau-Quesnel, Laval University, ping-pong matching scheme developed in Dzha- Zhuangzhuang Xi, Hye Joo Han, Purdue farov (2006, Journal of Mathematical Psychology, University, Claudette Fortin, Laval Univer- 50, 74-93). The statistics of the adjustment se- sity . Social networks in waking life are complex, ries have been found to be in a good compliance but tend to have regularities such as a power with the regular well-matchedness hypothesis. law degree distribution. An individuals memory for people and their relations presumably rep- Sunday, 10:00 (Session C) resents these regularities. When an individual Symmetry-Based Methodology is in REM sleep, primary sensory cortices are for Decision-Rule Identification blockaded, so there is little input from the out- in Same-Different Experiments. side world. It is reasonable to conclude that any Alexander Petrov, The Ohio State Uni- regularities that appear in dream content arise versity . The standard practice to reduce from corresponding regularities in the dreamers every same-different data set to two numbers memory. Social networks were constructed for (hits and false alarms) is wasteful because characters in dreams in a series from the same the response pattern to all four stimulus pairs dreamer. Considerable regularity appears; for carries information about the decision rule example, the largest eigenvalues of the adjacency adopted by the observer. We describe eight matrix follow a power law. rules organized in three families: differencing, Sunday, 12:10 (Session C) covert classification, and likelihood ratio. We Matching Regularity for 2D Shapes and prove that each family produces a characteristic Locations. Lacey Perry, Ehtibar Dzha- pattern of (in)equalities among the response farov, Purdue University . The general defini- probabilities. The basic idea of the proof is tion of well-matched regular stimulus spaces is that symmetric decision boundaries demarcate given in Dzhafarov & Dzhafarov (2010, Theoria, symmetric response regions, the perceptual 76, 25-53). When applied to pairwise presen- distributions have (mutually) symmetric prob- tation of stimuli in two fixed areas ability density functions, and the integrals of (e.g., one on the left, one on the right), the no- symmetric densities over symmetric regions are tion implies that (1) for every stimulus x one equal. On this basis, we propose two simple but can find a stimulus y in another observation area powerful qualitative tests for same-different data such that x and y match each other; and (2) analysis. Is the performance on stimulus pairs

40 AA and BB statistically indistinguishable? If Choice Reaction Time: An Information- not, differencing and likelihood-ratio strategies Discounting and Memory-Forgetting can be rejected. Is the performance on pairs Model. Yajun Mei, Georgia Institute of AB and BA indistinguishable? If not, covert Technology . In this talk, spurred by the classification can be rejected. We present two-choice experiments in which the subject algorithms (and Matlab software) for fitting does not know the time of occurrence of the two covert-classification models and illustrate signal, we extend Ratcliffs diffusion model to the new methodology in a the 2-CUSUM process model by emphasizing experiment on visual motion-direction discrimi- natural connections to the sequential change- nation. The standard assumption of symmetric point detection problems in statistics. Unlike decision criteria was violated. existing sequential sampling models which are based on Wiener, Ornstein-Uhlenbeck, or Poison Sunday, 9:20 (Session B) processes, the proposed 2-CUSUM process Adaptive Optimal Inference and Feed- model utilizes another family of widely used back for Numerical Representational stochastic processes: the regulated or reflected Change. Yun Tang, Christopher Young, Brownian motions. There are two main new Jay Myung, Mark Pitt, John Opfer, The features in the proposed 2-CUSUM process Ohio State University . We explore an adaptive model: information discounting in the sense of design optimization approach for optimizing ex- subtracting a non-negative term to Brownian perimental designs to discriminate mathematical motions (or Wiener processes) in the cumulative models of childrens number concepts. Studies information processes, and memory forgetting on the development of numerical representations in the sense that the decision makers will forget typically find that young children initially or ignore those “not so significant” information estimate numerical magnitudes to increase that is not consistent with their prior knowledge, logarithmically with actual value before later and will start afresh to cumulative information learning the decimal system (Siegler & Opfer until there is sufficient evident to prove that 2003). Recent evidence suggests that rapid and their prior knowledge information is incorrect. broad adoption of linear representations can be Surprisingly, Ratcliff’s diffusion model turns out induced in situ by providing feedback on a few to be a special case of the proposed 2-CUSUM key numbers (Opfer & Siegler, 2007). In this process model when the information discount- paper we develop and implement a Bayesian ing rate is zero. However, if the information inferential procedure to detect and facilitate discounting rate is positive, their properties representational change in numerical estima- can be completely different. For example, from tion. The proposed procedure of an adaptive the asymptotic viewpoint, the reaction time numerical experiment both infers a learner’s of Ratcliff’s diffusion model is distributed ac- representation and predicts the feedback that is cording to a mixture of Gaussian distributions, most likely to induce representational change. In whereas that of the proposed 2-CUSUM process the feedback session, in particular, we optimize model is a mixture of Gaussian and exponential the effectiveness of feedback stimuli by maxi- distributions. Directions for further research are mizing the discrepancy between the currently also discussed. inferred model of representation and the target model (e.g., straight linear line). We provide Sunday, 11:30 (Session C) an application of this procedure using simulated Probability judgments and partition de- subjects and demonstrate its effectiveness in pendence under sample space indetermi- inferring representational state and inducing nacy. Michael Smithson, The Australian Na- change. tional University, Jay Verkuilen, City Uni- Monday, 9:20 (Session C) versity of New York, Tsuyoshi Hatori, Tokyo The 2-CUSUM Process Model for Two- Institute of Technology, Michael Gurr, The

41 Australian National University . We extend the served quantity related to the syntactic structure study of partition priming effects to settings in of written numerals. We use the canonical en- which the sample space is ambiguous. Our ex- semble to state that the probability distribution perimental paradigm presents judges with k al- of the syntactic structure of the written numerals ternatives but no indication of whether the al- follows a Gibbs-Boltzmann distribution and to ternatives exhaust the space. A prime induces derive analytical expressions for thermodynamic some judges to interpret the space as having a quantities that describe the macroscopic state of k-fold partition whereas others interpret the par- the system. The empiric distributions for the tition as more than k-fold. The resulting sums syntactic structure of written numerals obtained of each judges probabilities are modeled by a fi- in the three sessions fit to Gibbs-Boltzmann dis- nite mixture of a beta distribution and one or tribution with coefficient of determination R- more degenerate distributions. This model han- Squared 0.9651, 0.9959 and 0.9807. We find for dles predictors in each of three submodels (loca- the three sessions that Gibbs entropy, tempera- tion and dispersion of the non-degenerate distri- ture and internal energy decrease over time while bution, and mixture composition). The model is Helmholtz free energy increases linearly. fitted to real data from experiments, using max- imum likelihood methods for independent obser- Sunday, 11:50 (Session B) vations and MCMC methods for dependent ob- Assessment Rate in Sequential Risky De- servations. cision Making. Avishai Wershbale, Timo- thy Pleskac, Michigan State University . Cur- Monday, 16:20 (Session C) rent cognitive models of sequential risky decision A statistical mechanics treatment of chil- making tasks generally assume the same response dren’s learning process to write Ara- process occurs at every opportunity (Wallsten et. bic numerals. Rafael Hurtado, Carlos al., 2005, Busemeyer & Stout, 2002). For ex- Quimbay, Universidad Nacional de Colombia, ample, during the Balloon Analogue Risk Task Mariela Orozco-Hormaza, Diego Guer- (BART), participants pump a computerized bal- rero, Universisdad del Valle, Alex Santos, loon and earn money for each successful pump, Universidad Nacional de Colombia . We use re- but gain no money if the balloon pops. The sults from cognitive psychology and statistical BARTs cognitive model posits a response pro- mechanics to develop a mathematical model for cess where at each and every pump opportu- childrens learning process of three-digit Arabic nity/decision point; participants account for the numerals. The experimental setup consisted in distance from a target and base their response asking 51 children in first grade of elementary (pump or stop) off of an assessment of this dis- school to write down in Arabic format a set of tance. Response time data, however, suggests 36 dictated numbers. The same set of numbers the existence of multiple response processes. A ordered randomly was dictated in three sessions slow goal directed response, consistent with the with two months of difference between consec- original model, but also a fast routinized re- utive sessions. Children production was classi- sponse selection process. Task exposure, as well fied from the syntactic structure of written nu- as distance from the targeted pump, affects the merals by using a cognitive model based on a rate at which assessments are made, with task ex- semantic-lexical internal number representation. posure decreasing the rate of slow goal-directed This model proposes sum and product structures responses and target distance increasing the rate. as the main relationships between numerical con- These results motivate the development of alter- cepts (Power, R. J. D., & Dal Martello, M. F. native process models of the BART. Maximum (1990). The dictation of Italian numerals. Lan- likelihood model comparisons suggest the best guage and Cognitive Processes, 5, 237-254). We model assumes a dynamic trial dependant assess- model the social system as constituted by the set ment rate. More generally, the idea that not ev- of written numerals and with an energy-like con- ery risky decision that is made is necessarily pre-

42 ceded by a goal directed assessment is similar to pattern of association-memory impairments in vicarious trial and error rates that occur in rats medial temporal lobe amnesia. during maze learning (Tolman, 1948). The mul- Tuesday, 11:30 (Session B) tiple response processes may also help explain Binarizing Multi-link Multinomial known differences in BART performance that Processing Tree (MMPT) models. were found between conduct disorder adolescents Xiangen Hu, The University of Memphis, and their control group counterparts (Crowley et. William Batchelder, University of Califor- al., 2006). nia, Irvine . Multi-link multinomial processing tree (MMPT) models parameterize rooted trees Tuesday, 15:20 (Session B) by assigning manifest categories to the leaves The convolution operation for associative and parameter vectors to the non-terminal nodes memory: (nearly) found in the brain. of the tree. In particular if a non-terminal node Jeremy Caplan, University of Alberta .Two has u>2 children, then the parameter vector mathematical frameworks can explain a broad assigned to the node will have u component range of human memory behaviour: Matrix mod- parameters that are non-negative and add to els (Anderson, 1970) use matrix multiplication to one. The special case where every non-terminal store associations, whereas “holographic” mod- node has exactly two children leads to the class els (Westlake, 1970; Borsellino and Poggio, 1971) of binary MPT (BMPT) models, which is well use convolution. Matrix multiplication has been understood mathematically and statistically. thought of as neurally plausible due to its simi- This paper will present some formal results that larity to learning in connectionist networks and connect MMPT and BMPT models. We will being viewed as plausibly carried out by the cir- show that for any MMPT there is an associated cuitry within the hippocampus; this has been set of BMPT models each of which is statistically used to push convolution-based models on the equivalent to the MMPT model. We call this back burner as neurally implausible (Pike, 1984). collection of such models a ‘binarization’ of the A recent surprising finding - “grid cells” in medial MMPT model. An obvious advantage of a bina- entorhinal cortex (Hafting et al., 2005), a ma- rization is that it allows us to utilized existing jor input to the hippocampus - appear to form theory and implementations for BMPT models. a 2D spatial Fourier basis set for a place rep- We will focus on the following mathematical resentation thought to be held within the hip- and statistical questions regarding binarization: pocampus (Solstad et al., 2006). Fourier trans- 1) How to construct a general binarization algo- form is a well known mathematical shortcut to rithm? 2) How to use binarization to improve compute convolutions in O(n log(n)) (Brigham, statistical analysis? For example, given a set 1974). I suggest that the grid-cell finding tells of parameters of interest in an MMPT model, us that the hippocampus carries out spatial, as how to obtain a binarization that contains well as non-spatial associative learning via con- only the selected parameters (hence ignoring volution (in addition to - or conceivably instead other parameters), 3) How to understand the of - matrix multiplication), supporting the neural mathematical structure of the set of ‘minimal plausibility of convolution-based memory models binarizations’ of a given MMPT with respect to and potentially explaining how the hippocampus a set of parameters. implements associative symmetry (Bunsey and Eichenbaum, 1996), a property of convolution- Tuesday, 9:40 (Session C) based models wherein there is no directionality Comparing Three Different Equa- stored within associations (Asch and Ebenholtz, tions Representing Utility for a 1962). This insight suggests that hippocampal Single Reinforcement Schedule. modelers may need to consider convolution as Robin Gane-McCalla, Dare Institute . a computational function of the hippocampal- Counting trial time and applying p to a sequence entorhinal pathway, and may help explain the of trials, produces a game theoretical/economic

43 equation of utility (Von Neumann & Morgen- with the assumptions of the diffusion model. stern, 1944). Commons, Woodford & Trudeau Second, we show by simulation that model (1991) proposed a second way to examine equivalence only holds for boundary separation utility, which will be called the Quantitative values unlikely to occur in practice. Analysis of Behavior Equation. It is applied to Tuesday, 16:20 (Session A) a single t reinforcement schedule (Schoenfeld & A Dynamic Model for the Emergence of Cole, 1972). This latter schedule and equation Two-sided Social Conflicts. Seth Marvel, application may be considered as intermediate Jon Kleinberg, Robert Kleinberg, Steven between the Game Theoretical/Economic Model Strogatz, Cornell . When seized by certain and Herrnstein’s (1970) Matching Law applied types of social conflict, human communities show to variable interval schedules. The matching a remarkable tendency to divide into two camps. law is: R1 = kr1/r1 + re , where R1 = rate of This phenomenon can be observed in the compe- response, r1 is the rate of reinforcement, and re tition among political parties, standard-setting is reinforcement besides r1. We assume rate R1 in the marketplace, and the outbreak of inter- is a behavioral measure of utility. We show all national war. We consider here a nonlinear sys- three equations are essentially the same when tem dX/dt = X2 (where X is a square matrix) applied to the t-schedule. For our t-schedule, that exhibits this general behavior—for generic vi = p(vi) ∗ vd , where vi = effective or perceived initial conditions, it sorts itself into two factions reinforcer value at time i, vd· = value with a over time. This system also has a natural inter- given delay Under the Commons et al equation, pretation in terms of human psychology: its form given a sample with 4 cycles of length 2 secs, can be qualitatively justified using social balance a pigeon receives a reinforcer with p = .25, or theory. On the empirical end, we find that the p = .75, the p being similar to the probability of system gives results consistent with various pre- a reinforcer in the traditional utility case. With viously studied cases of social fission, for exam- the addition of trial time, and multiple trials, ple Zachary’s karate club division and the World the utility equation was similar to the Commons War II Allies/Axis division. On the theoretical et al value equation because the values add after end, we prove a collection of mathematical re- being discounted appropriate to their time away sults that together help explain how the two-way from choice. Likewise the matching law equation splitting process occurs over the course of the and the Commons et al equation support that idealized dynamics. Included in these results is a the key variable is the rate of reinforcement. procedure for predicting the membership of the two sides without integrating dX/dt = X2, and a Sunday, 16:20 (Session C) characterization of the expected time-dependent Model Equivalence between the evolution of the system in the large-population Diffusion Model and the LCA. limit. Don van Ravenzwaaij, Han van der Maas, Eric-Jan Wagenmakers, University of Amsterdam . In their influential Psychological ACCEPTED POSTERS Review paper, Bogacz, Brown, Moehlis, Holmes, and Cohen (2006) used phase planes to inves- Sunday, 17:00 (Poster Session) tigate the conditions under which the diffusion The Predictability and Abstractness of model (Ratcliff, 1978) and the leaky, competing Language: A Study in Understanding and accumulator model (Usher & McClelland, 2001) Usage of the English Language through may be mathematically equivalent. In this Probabilistic Modeling and Frequency. study, we extend their findings and identify two Revanth Kosaraju, The Harker school . Ac- challenges. First, the model equivalence implies counts of language acquisition differ significantly that boundary separation in the diffusion model in their treatment of the role of prediction in lan- changes from trial to trial, which is inconsistent guage learning. In particular, nativist accounts

44 posit that probabilistic learning about words and Predicting behavior from genetics with word sequences has little to do with how chil- correspondence analysis. Derek Beaton, dren come to use language. We examined the Herve´ Abdi, The University of Texas at Dallas accuracy of this claim by testing whether distri- . A recent important trend in cognitive neuro- butional probabilities and frequency contributed science studies integrate behavioral and genomic to how well 3-4 year olds were able to repeat data. For genomics, a popular approach identi- simple word chunks. Corresponding chunks were fies, in DNA, loci of variation, which are called the same length, expressed similar content, and single nucleotide polymorphisms (SNPs). Tra- were all grammatically acceptable, yet the results ditionally, in studies, be- of our study showed marked differences in per- havioral data are collected with questionnaires formance when overall distributional frequency and psychometric instruments. Both behavioral varied. We found that a distributional model and genomic data possess comparable structural of language predicted our empirical findings bet- properties. First, these data are organized by ter than a number of other models, replicating blocks, (e.g., chromosomes for a genome; mem- earlier findings and showing that children attend ory or perception for behavior). Second, these to distributional probabilities in an adult corpus. data often have (i.e. subjects) that This suggested that language is more prediction- can be grouped based on criteria such as clinical and-error based, rather than on abstract rules diagnosis or familial relationship. Interestingly, which nativist camps suggest. each of these structures (genomics and behav- ior) can be analyzed with Multi-block Correspon- Sunday, 17:00 (Poster Session) dence Analysis (MuCA) and Multi-block Dis- An ANS-Based Model of Children’s Small- criminant Correspondence Analysis (MuDiCA) Set Size Judgments. James Negen, Bar- that are recently developed variations of cor- bara Sarnecka, University of California, respondence analysis (CA), which in turn, is Irvine . Before learning how to use counting effec- an eigen-decomposition technique for analyzing tively, young children first learn the meaning of qualitative data. MuCA projects blocks or ta- the first four number words (e.g., Wynn, 1992). bles (i.e. genomic and behavioral data) in the We present a model of how children in this devel- same factor space, while MuDiCA creates a fac- opmental period respond when you ask them to tor space to discriminate groups (i.e. clinical estimate how many items are in a picture. This groups or families). Additionally, we want to model assumes that children use the Approxi- reveal common and predictive information be- mate Number System (ANS), which is a cogni- tween genomics and behavior. This is achieved tive system that provides a non-verbal estimate by adapting Partial Least Squares Regression for of the number of items in a given stimulus (for CA. Furthermore, MuCA and MuDiCA incor- review, see Feigenson, Dehaene, & Spelke, 2004). porate an inferential component, through cross- The model infers the distribution of latent con- validation techniques such as jackknifing, boot- tinuous activation intensities in the child’s ANS strapping and permutation tests, which can be from the distribution of discrete verbal responses. used to develop and refine hypotheses about We present several ways in which this provides cognitive, behavioral and genetic development. a good fit to data; as our focus, we show that We present CA, MuCA and MuDiCA and illus- scalar variability (the standard deviation of an- trate them with an example integrating genomics swers growing linearly with the mean), which is (SNPs) and behavioral data (neuropsychological a well-known signature of the ANS, can be in- and neuroimaging). ferred from a large corpus of child data. This evidence speaks against the claim that the ANS Sunday, 17:00 (Poster Session) is not used in early number-word learning (Le Mathematical Modeling of Memory Evolu- Corre & Carey, 2007). tion - Pure Decay. Eva Cadez, Evan Heit, Sunday, 17:00 (Poster Session) University of California, Merced, Vladimir

45 Cadez, Astronomical Observatory Belgrade, memory models be derived further, in order to Serbia . Memory processes possess properties of identify loci within the major models where item- a physical dynamic system and can be qualita- level versus association-level effects could mate- tively described by a suitable set of differential rialize. Here we present a systematic approach equations. In this work, we analyze analytical re- to modeling item- versus association-memory ef- sults of a specific mathematical model that sim- fects in PA learning, with a specific focus on com- ulates a time dependent processing of a mem- paring the Matrix Model and TODAM, and us- ory pattern of a person. The memorized pattern ing BSB to model redintegration effects. These evolves in time which can be qualitatively de- derivations and simulations suggest a range of scribed by a single first order differential equation model-relevant effects that item properties could typical of non-uniform damping and attenuation have, some of which are suggestive of existing processes found in physics. Such an equation is empirical results and others of which may fore- solved analytically for various mathematical ex- shadow future findings. pressions for the initial pattern, simulating lists Sunday, 17:00 (Poster Session) of objects memorized i.e. learned, and for model What Can the Diffusion Model Tell functions for the attenuation rates. The atten- Us About Prospective Memory? uation rate modifies decay of memory without Sebastian Horn, Ute Bayen, Heinrich- rehearsal. We show that it can simulate the ob- Heine-Universitaet Duesseldorf , Rebekah ject and time dependent individual attention or Smith, The University of Texas at San Antonio an influence of task instructions warning about . Prospective memory (PM) involves remember- false memories in paradigms involving process ing to perform intended actions in the future of memorizing patterns. The obtained solutions (i.e., “remembering to remember”) and often are discussed and plotted in a normalized form occurs in the midst of other ongoing tasks. that shows a time variation of relative ordering Ongoing task performance can be negatively of memorized objects, which may represent their affected by an additional PM task. This cost- subjective significance. The significance is sim- or interference effect of PM is observed on non- ulated and discussed in terms of findings about target trials, even before the first opportunity familiarity and temporal phenomena in list recall to perform the intended actions. Surprisingly experiments. Current literature is not clear on little is known about the specific reasons for roles of pure decay of memory partially because these cost effects. In addition, extant analyses it is extremely hard to isolate it from processes with ongoing lexical decision tasks were usually which are postulated to have similar role in for- restricted to RTs of correct word trials. We getting. We hope this view of its dynamics, with extend this approach by applying Ratcliff’s simulations of existing data, may help clarify the diffusion model in the PM paradigm to examine concept and aid research on decay. the processing components underlying cost effects and their relation to PM performance. Sunday, 17:00 (Poster Session) The model-based results suggest that non- Model-mechanisms for effects of item- decisional RT components and criterion shifts on association-memory in paired-associate explain a substantial proportion of the change learning. Christopher Madan, Jeremy Ca- in ongoing task performance and can thereby plan, University of Alberta . Paired-associate inform current theories of PM. (PA) learning paradigms are extremely common in memory research; however, memory behaviour Sunday, 17:00 (Poster Session) in these paradigms relies on both memory for The Wisdom of the Crowd Playing the items and for their pairings. Several mathemati- Price is Right. Michael Lee, Shunan cal modeling frameworks have been applied suc- Zhang, Jenny Shi, University of California cessfully to PA empirical phenomena. However, Irvine . In the “Price is Right” game show, play- our new behavioural results demand that these ers compete to win a prize, by placing bids on its

46 price. We ask whether it is possible to achieve a stimuli in terms of a general framework that rep- Wisdom of the Crowds effect, by combining the resents categorization processes by humans and bids to produce an aggregate price estimate that machines, and will then propose a mathematical is superior to the estimates of individual players. formulation of the phenomenon. This general ap- Using data from the actual game show, we show proach will be illustrated using several paradigms that a Wisdom of Crowds is possible, especially in human and machine classification. by using models of the decision-making processes Sunday, 17:00 (Poster Session) involved in bidding. The key insight is that, Instruction determines the error pattern because of the competitive nature of the game, in relative order judgments. Yang Liu, what people bid is not necessarily the same as Michelle Chan, Jeremy Caplan, Univer- what they know, and better estimates are formed sity of Alberta . Judging the relative order of aggregating latent knowledge than observed bids. events is a core function of human memory. In We discuss what our results say about studying short, (sub-span) consonant lists with immediate competitive group decision-making. We finish by judgments of relative recency (JOR), instruction arguing our results also highlight the fundamen- wording (“which item was presented earlier?” tal role models of cognition and decision-making versus “which item was presented later?”) ap- should play in solving real-world problems typi- peared to reverse memory search direction (Chan cally approached only from the non-psychological et al., 2009). We wondered whether instruction perspectives offered by statistics and machine wording would have an analogous influence on learning. the JOR judgment in supra-span lists. However, supra-span lists are more complicated; both a re- Sunday, 17:00 (Poster Session) cency effect and a distance effect are typically Detection of Rare Events and Categoriza- observed (e.g., Yntema & Trask, 1963). Partic- tion. Max Quinn, Holly Jimison, OHSU , ipants performed JORs on sub-span and supra- Daphna Weinshall, Hebrew University of span consonant lists (Experiment 1; LL = 4 ,8) Jerusalem, Frank Ohl, Leibniz Institute for or a range of supra-span noun lists (Experiment Neurobiology, Hynek Hermansky, Dept. Elec- 2; LL = 4, 6, 8, 10). Linear mixed-effects analysis trical and Computer Engineerin, Misha Pavel, revealed a similar instruction effect on supraspan OHSU . The detection of rare and novel events lists as for subspan lists, consisting of an increase or stimuli, at the crux of many intelligent sys- in response time (RT) performance with increas- tems, has a long history in psychology, pattern ing target position for the “Later” instruction recognition and machine learning. There are sev- and a decrease in RT performance with target po- eral interpretations of these notions that have sition for the “Earlier” instruction. Importantly, surfaced recently in and instruction influenced the error-rate pattern for that have increasingly been incorporated into en- both subspan and supraspan JOR data. A ver- gineering systems. In this paper we will first re- sion of SIMPLE model designed to explain serial view and contrast mathematical representations recall data with the addition of a serial-position of three well-established versions of novelty de- gradient (Brown et al. 2007) which we view as tection: oddball detection, salience and surprise. an implementation of an attentional bias expla- Each will be described within a common frame- nation of the data fits the chief qualitative fea- work involving probability distributions over in- tures of the data including the difference between put features, utility on outcomes, and low level instructions. Our findings and model fit suggest models. We will then propose an interpretation that the JOR task may be understood as a two- of novelty in terms of incongruence. An example dimensional judgement, the second dimension re- of an incongruent stimulus is a clearly heard but flecting a congruency-based bias across position. unknown word, which is related to the impor- tant task of out-of-vocabulary word detection. We will describe the detection of incongruent Sunday, 17:00 (Poster Session)

47 Assessing psychometric functions and to making successful empirical predictions about thresholds using unforced-choice adaptive the learning difficulty ordering of key Boolean methods. Kai-Chun Chang, Yung-Fong categories involving discrete dimensions, we show Hsu, National Taiwan U. . One of the long- here that the CIM is highly flexible in that it standing interests in is the study can be easily extended to continuous cases by of thresholds for discriminability. While thresh- generalizing the DNF (disjunctive normal form) old theories (e.g., Luce, 1963) embrace the con- description of the categorical stimulus to allow cept of (discrete-state) thresholds, signal detec- for continuous values in the real number interval tion theory (Green & Swets, 1966) discounts such [0, 1]. In this theoretical note, we use the cDNF a concept. Until very recently, the meaning of (continuous DNF) of a concept function as de- thresholds was still not fully understood (see fined in CFMS multi-valued logic (Cattaneo et Rouder & Morey, 2009). We are concerned with al., 1993). This note also serves as an introduc- the practical issue of estimating threshold val- tion to the CIM. ues. In doing so, we find it necessary to clarify the concept germane to the psychometric func- Sunday, 17:00 (Poster Session) tion, which is customarily constructed using psy- Applications of Shannon’s En- chophysical methods with a binary-response for- tropy Theorem to Psychology. mat. We then argue that confidence responses Koralur Muniyappa Mahendra, Airvana are also important for the construction of psycho- Networks India Pvt Ltd, India . Applications metric functions and the estimate of thresholds. of Shannon’s Entropy Theorem to Psychol- We outline an algorithm to incorporate confi- ogy. Background Motive: Shannon’s Entropy dence responses into a family of (non-parametric) Theorem is proposed to measure how much adaptive methods for threshold estimation. Typ- uncertainty or information is associated with ical examples are the three-category weighted up- the occurrence of a probabilistic event. Since down (WUD) method proposed by Kaernbach many of the stimulus (especially emotional (2001) and the four-category WUD suggested by stimuli) occurrence events are probabilistic in Klein (2001). We argue that such a rating-scale nature or are not certain events which make the response algorithm can be further extended to Shannon’s Entropy Theorem easily applicable the continuous version, for which a response al- to general-behavior prediction. Methodol- gorithm called ”visual analogue scale” can be ap- ogy/Model: Shannon’s Entropy Theorem plied. states that the entropy H of an experiment with n number of outcomes is given by the Sunday, 17:00 (Poster Session) formula H = K P[p(i)log(1/p(i)]; i=1ton; An Extension of the Categorical In- K=constant. From above we obtain the expres- variance Model to Continuous Domains. sion for the amount of uncertainty (I) associated Ronaldo Vigo, Ohio University, John Kr- with the occurrence of each outcome of the uschke, Indiana University . The categorical in- experiment as below I = Klog(1/p); The term variance model (CIM) developed by Vigo (2008, I corresponds to probabilistic-component of the 2009a, 2009b, 2009c), has been successful in pre- response given by the brain to the stimulus with dicting the degree of subjective difficulty expe- probability of occurrence p. Hence the total rienced by humans when learning concepts de- response (R) is equal to the sum of probabilistic- fined by Boolean rules. The CIM introduces the component (I) and normal-component (N). concept of a logical manifold for measuring the R = I + N; N=normal-response to the stimulus degree of structural homogeneity of a Boolean treating it as certain event. The above concep- category. The CIM posits that the structural tualizations been utilized to explain below and complexity of a Boolean category is inversely many more. 1) Maximization Principle: How to proportional to its degree of invariance and di- optimize the efficiency of the all human/animal rectly proportional to its cardinality. In addition activity together minimizing the cost involved.

48 2) Defining abstract terminology in field of Kornblum, Requin, Whipple, & Stevens (1999) psychology. 3) Explain Audience psychology manipulated the temporal delay (stimulus onset when he listens to a song or story/incident or asynchrony or SOA) that separates the presenta- watches a movie 4) Proving “There is no end tion of irrelevant from relevant information. The to Human desires”. 5) Explains important and results showed that the time course of the inter- standard emotional/social behavior. ference effect differs between the Stroop and the Simon task. These and other results appear to Sunday, 17:00 (Poster Session) put great constraints on how mathematical mod- Cognitive Modeling Repository. els account for interference in the Stroop and Jay Myung, Mark Pitt, Daniel Cav- Simon task. For instance, Stafford & Gurney agnaro, Ohio State University . Quantitative (2007) pointed out that the diffusion model can- modeling has contributed substantially to the not handle empirical data from interference tasks advancement of the cognitive sciences. Papers when SOA is very long. Here we seek to repli- introducing and testing cognitive models regu- cate the findings of Kornblum et al. (1999) while larly appear in the top journals. The growth and including a very long SOA condition. Moreover, success of cognitive modeling demonstrate why we have implemented Kornblum’s computational modeling itself should be a primary quantitative model and compare its ability to describe the method in the researcher’s toolbox. Yet this data to different extensions of the drift diffusion method of scientific investigation remains under- model. We conclude that developing and test- utilized by the research community at large, in ing mathematical models to conflict data can ad- part because of the hassles in obtaining data sets vance our understanding of the underlying mech- to model and the difficulties in implementing anisms of interference. models. The goal of this project is to assist scientists in their cognitive modeling efforts by Sunday, 17:00 (Poster Session) creating an online repository containing data Bow Effects in Absolute Identification. sets that can be modeled and the cognitive Pennie Dodds, Chris Donkin, Indiana Uni- models themselves. The current state of the versity, Scott Brown, Andrew Heathcote, project and future plans will be presented. The University of Newcastle . The bow effect is Prior to the conference, the work-in-progress one of the key phenomena observed within ab- repository website (www.cmr.osu.edu) will be solute identification. Normally characterised by opened up to conference attendees for their poor performance for stimuli in the centre of the comments and feedback. The project is funded set relative to the surrounding stimuli, the bow by the Mathematical Modeling of Cognition effect is so named as it forms a U shaped curve and Decision program of the Air Force Office of when plotting stimuli against performance. We Scientific Research. demonstrate that this bow effect can be distorted Sunday, 17:00 (Poster Session) when using a within-subjects manipulation of the Modeling the Time Course of Interfer- number of stimuli making up the set, leading to ence. Annika Boldt, Humboldt-Universit¨atzu an atypical W-shaped bow effect. This W ef- Berlin, Roger Ratcliff, Ohio State Univer- fect is characterised by improved performance for sity, Eric-Jan Wagenmakers, University of stimuli in the middle of the set relative to their Amsterdam . In the Stroop task and the Simon surrounding stimuli. We show that this effect task, irrelevant information interferes with the is caused by the additional presentations of cen- processing of the relevant stimulus attributes, tre stimuli which arise naturally from a within- causing an increase in response times and error subjects manipulation of set size, and that the rates. One popular method for studying the tem- effect is not simply the result of response biases. poral dynamics of these interference effects is to These findings strongly suggest that a within manipulate the extent to which the irrelevant in- subjects manipulation of set size is inappropriate formation gets a head start. In a seminal study, for absolute identification tasks, and that theo-

49 ries of absolute identification must account for ysis of the response trajectories (mousepaths) stimulus-specific, long-term improvement due to from participants in risky decision making tasks additional presentation frequency. while moving from a neutral region to screen re- gions associated with the selection of various op- Sunday, 17:00 (Poster Session) tions. Even though participants were unaware Functional Principal Component Analy- that their mouse movements were being recorded sis and the Capacity Coefficient. Devin (they were not explicitly tracking changes in pref- Burns, Joseph Houpt, James Townsend, erence), we see evidence of the implicit, com- Indiana University . The capacity coefficient is petitive “pull” from foregone choice alternatives a well established measure of the efficiency of in the mousepaths towards chosen alternatives. processing combined sources of information. It We also find dependencies in the nature of these has been applied to measure cognitive processes mousepaths on the risk of the chosen alternative, ranging from audio-visual integration to face per- as well as changes occurring over repeated trials. ception. Recently, the capacity coefficient has We suggest how these novel metrics can provide also been applied in various clinical situations. new insight into the cognitive processes underly- Typical clinical analysis, such as structural equa- ing decision making. tion modeling, use scalar values or vectors with limited length as input. We explored the use of Sunday, 17:00 (Poster Session) functional principal component analysis (fPCA) Assessing perceptual correlations in the to allow researchers to describe the capacity co- uncertainty paradigm: Exploring predic- efficient, a continuous function of time, with a tions of decision models within the mul- small set of discrete values. The fPCA approach tidimensional signal detection framework. was compared with two simple alternatives for re- Robin Thomas, Ryan Brunton, Miami Uni- ducing the capacity coefficient to a single value. versity . The uncertainty paradigm has been used We found that fPCA captured the major trends in vision research to evaluate whether stimulus in the data more effectively than other methods. components are processed independently or not. In typical applications of this paradigm, a mut- Sunday, 17:00 (Poster Session) licomponent stimulus differs in one of its com- Response dynamics and the time course ponents from a standard value and the observer of risky decision making. Joseph Johnson, needs to decide if the change is an increment or Gregory Koop, Evan Bristow, Miami Uni- decrement. In the certainty condition, the ob- versity of Ohio . Process-tracing techniques such server knows which component will contain the as mouse- and eye-tracking have become popular change; in the uncertainty condition, the com- for drawing inferences about the information ac- ponent that differs from standard is unknown. quisition process in decision-making–but not nec- Performance across the two conditions can be essarily the dynamic mapping of that informa- compared to that which is predicted by inde- tion onto a response (utilization). Rather, even pendence of components. We derive predictions these very studies continue to treat responses dis- for observer sensitivity in the uncertainty con- cretely. In contrast, we propose that closer exam- dition and a relative measure, root-mean-square ination of the time course of the response process (RMS) that incorporates both uncertainty and can yield new insights. The current work draws certainty performance for three major decision upon a growing body of research in cognitive sci- models when using a signal detection theory ence focusing on “action dynamics” (Dale, Ke- framework stimulus components are perceptually hoe, & Spivey, 2007; McKinstry, Dale, & Spivey, correlated: a distance-classifier, the optimal de- 2008; Spivey & Dale, 2006). This research has cision model (ideal observer), and a decisionally shown quite clearly that mental processing dy- separable (independent decisions) strategy. Data namics can be revealed in the associated con- from an identification task and the uncertainty tinuous motor response. We present an anal- paradigm (using the same stimulus components)

50 are analyzed from the perspective of these the- cal stress-coping systems. oretical results and model-fitting to assess the Sunday, 17:00 (Poster Session) validity of the predictions. How is Story Line Dependent Context Represented in the Brain? An inves- Sunday, 17:00 (Poster Session) tigation using EEG and MUBBADA. “Hot Cognition” is Not Beyond the Reach Mary Kathryn Reagor, Anjali Krishnan, of Formal Models: Higher-Dimensional James Jerger, Joseph Dunlop, Herve´ Abdi, Nonlinear Bifurcations Depict Psycholog- University of Texas at Dallas . In everyday life, ical Stress and Coping. Lawrence Levy, we have many conversations (narratives) with University of Western Ontario, Weiguang different people at different times and places. Yao, Med. Physics, Sudbury Hosp./Laurentian How do we associate the information from these U , George McGuire, University College of narratives? We must package it somehow be- Fraser Valley, Daniel Vollick, U. British cause usually, once we remember one or two de- Columbia/Okanogan, Richard Neufeld, U. tails, the rest seems to come flooding back into Western Ontario . A long-held canon of psy- memory. Previous studies have shown that se- chological stress and coping is that they are mantic relations can be used to prime a tar- quintessentially dynamical and transactional, en- get word, but what about context? Here, we tailing reciprocal control and feedback mech- were interested in whether context alone facili- anisms. The mechanisms incorporate coping tates word recognition. EEG was recorded while activity, stressor level, and collateral variables participants listened to stories and were then pre- such as subjective appraisal of coping effective- sented a series of words. We wanted to know if a ness.Focusing on a prominent form of coping word from one of the stories primed the recogni- known as “decisional control” – positioning one- tion of the next word if it was also from that self in a multifaceted stressing situation so as to story (positive priming). Conversely, negative minimize the probability of an untoward event priming might occur if the next word was from – established observations on stress-coping rela- another story, making it more difficult to rec- tions are expressed in the form of a 6-dimensional ognize. To analyze the EEG data, we used a nonlinear dynamical equation. Analytical and new multivariate method, Multi-Block Barycen- numerical analyses disclose rich dynamical prop- tric Discriminant Analysis (MUBBADA). MUB- erties of the resulting system, with bifurcations BADA has been designed to handle data with from stable and unstable fixed-point equilibria many more variables than observations, and is (higher-dimensional saddle node), through limit- therefore well suited for EEG. The goal of MUB- cycle repetitions, to chaos. Each attracting pat- BADA is to predict group membership (condi- tern corresponds with meaningful stress-coping tion) of observations and determine the discrimi- relations. Stable equilibria take the form of nant contribution of blocks of variables (e.g. elec- elevated efficiency of decisional-control related trodes). MUBBADA gives factorial maps where cognition, combined with correspondingly low the observations and categories are represented prevailing stressor levels, the opposite combina- as points with confidence intervals. Using MUB- tion characterizing unstable equilibria. Substan- BADA we can determine: 1) whether robust tively anchored local analysis discloses a Hopf- differences exist between negative and positive bifurcation structure that points to potential priming conditions, and 2) which electrodes con- control of the parameter boundary where the sys- tribute to the discrimination. tems’ fixed-point equilibria lose or re-attain sta- bility. Results inaugurate formal-model support Sunday, 17:00 (Poster Session) for a pivotal assumption about this domain of Wisdom of the Crowds Effects in Pre- human behavior, that intricate nonlinear dynam- diction of Future Outcomes; A case ics embodied in theoretical physical systems in study using Fantasy Football Rankings. principle extend to higher-dimensional theoreti- Timothy Rubin, Michael Lee, Mindy DeY-

51 oung, Mark Steyvers, UC Irvine . There ex- plains recognition data. The latter assumes that ists a large literature demonstrating the “Wis- recognition decisions are based on recollection dom of the Crowd” effect in aggregating across and familiarity, while the former assumes deci- large numbers of individual estimates. Broadly, sions are based on a single dimension of memory the explanation underlying this phenomenon is strength. Several authors (e.g., Pratte, Rouder, that individuals have access to some representa- & Morey, 2010) have been using hierarchical tion of the true value of what they are predicting, models, fit by Markov chain Monte Carlo meth- and that averaging across individual’s estimates ods, to explore sources of variance in recogni- will remove the independent errors that are in- tion tasks. However, no one has yet addressed troduced when generating each prediction. We the issue of model identifiability and flexibility consider an alternative situation, in which indi- using these methods. We examined the extent viduals must make predictions for future events. to which single and dual-process models could Specifically, we evaluate whether we can gen- be distinguished in a series of simulations and erate accurate predictions for the performance fits of Bayesian hierarchical models. Our results of individual players in the National Football have implications for the kind of data that can League (NFL) by aggregating across many in- discriminate between the models and the practi- dividual predictions. We show that these aggre- cality of using ROC curves as a basis for model gate predictions perform among the best individ- selection. ual human predictions, particularly when we con- sider performance across multiple sets of rank- ings. This research provides an interesting alter- native to traditional Wisdom of the Crowd ex- periments, because there exists no ground truth at the time of prediction. An additional contri- bution of this work is that is illustrates how to extend models for aggregating individual rank- ings under conditions in which data is missing not at random (MNAR).

Sunday, 17:00 (Poster Session) Bayesian Hierarchical Model Evaluation for the Receiver Operating Characteristic in Recognition Memory. Kenneth Olson, Trisha Van Zandt, The Ohio State University . One of the major tools for analyzing recogni- tion memory is the receiver operating character- istic (ROC), which is a curve giving the proba- bility of a hit (correct “old” response) as a func- tion of the probability of a false alarm (incorrect “old” response). Signal-detection theory (Green and Swets, 1966) has been widely endorsed, not only because the shape of the ROC curves it pro- duces are almost always consistent with data, but also because it provides a useful framework for describing the old/new decision. Recent de- bate (e.g., Wixted, 2007) has questioned whether the unequal-variance signal-detection model or a dual-process model (Yonelinas, 1994) better ex-

52 Author Index

Abdi, Herv´e [email protected], 10, 19, 20, [email protected], 26, 29, 45, 51 36 Addis, Kelly Brown, Scott [email protected],35 [email protected], 15, 49 Ambert, Kyle Brunton, Ryan [email protected],25 [email protected],50 Anders, Royce Burns, Devin [email protected],27 [email protected],50 Anderson, James Busemeyer, Jerome [email protected],34 [email protected], 32, 33 Andrews, Mark [email protected], 3, 29 Cadez,Eva Anwander, Alfred [email protected],45 [email protected],19 Cadez, Vladimir Aue, William [email protected],45 [email protected],6 Camilleri, Adrian Austerweil, Joseph [email protected],15 [email protected],37 Caplan, Jeremy [email protected], 43, 46, 47 Balci, Fuat Carney, Laurel [email protected],35 [email protected],9 Batchelder, William Catrambone, Joseph [email protected], 11, 16, 17, 27, 37, 43 [email protected], 13, 28 Bayen, Ute Cavagnaro, Daniel [email protected],46 [email protected], 21, 49 Beaton, Derek Chan, Michelle [email protected],45 [email protected],47 Belkin, Mikhail Chang, Kai-Chun [email protected],8 [email protected],48 Bogacz, Rafal Chater, Nick [email protected],19 [email protected],30 Boldt, Annika Chechile, Richard [email protected],49 [email protected],22 Borsboom, Denny Cheng, Chung-Ping [email protected],31 [email protected],28 Bracis, Chloe Cho, Min-Hyung [email protected],34 ,13 Bristow, Evan Cohen, Andrew [email protected],50 [email protected],33 Brown, Scott Commons, Michael

53 [email protected],21 Freestone, David Commons-Miller, Nicholas [email protected],35 [email protected],21 Corter, James Gane-McCalla, Robin [email protected],38 [email protected], 21, 43 Cousineau, Denis Gluck, Kevin [email protected],20 [email protected],35 Cox, Greg Golden, Richard [email protected],26 [email protected],16 Criss,Amy Goldstone, Rob [email protected],6,9 [email protected],12 Goldwater, Sharon Darlow, Adam [email protected],4 adam [email protected],26 Gonzalez, Richard Davelaar, Eddy [email protected],21 [email protected],5 Goodman, Noah Davis-Stober, Clintin [email protected],3 [email protected], 18, 22 Goodwin, Andrew DeCarlo, Lawrence [email protected],34 [email protected],19 Greene, Kshanti Dennis, Simon [email protected],12 [email protected],8 Griffiths, Thomas Denton, Stephen tom griffi[email protected], 2, 4, 37 [email protected],23 Guerrero, Diego DeYoung, Mindy [email protected],42 [email protected],51 Gureckis,Todd Dodds, Pennie [email protected],39 [email protected], 15, 49 Gurr, Michael Donkin, Chris [email protected],41 [email protected], 20, 49 Dunlop, Joseph Hagler, Stuart [email protected], 26, 51 [email protected],23 Dutilh, Gilles Han,HyeJoo [email protected],10 [email protected],40 Dzhafarov, Ehtibar Hatori, Tsuyoshi [email protected], 22, 40 [email protected],41 Hawkins, Guy Feldman, Naomi [email protected], 15, 36 naomi [email protected],4 Heathcote, Andrew Fields, Jerad [email protected], 20, 25, [email protected],39 33, 49 Forstmann, Birte Heit, Evan [email protected],19 [email protected],45 Fortin, Claudette Heller, Katherine [email protected],40 [email protected],30 Frank, Michael Hemmer, Pernille [email protected],4 [email protected],30

54 Henley, Steven Jones, Matt [email protected],16 [email protected],28 Hermansky, Hynek Jordan, Mike [email protected],47 [email protected],2 Herzog, Stefan [email protected],35 Kachergis, George Hidaka, Shohei [email protected],36 [email protected],32 Kahana, Michael Hills, Thomas [email protected], 8, 15 [email protected],31 Kashner, Michael Ho, Tiffany [email protected],16 [email protected],10 Kaznatcheev, Artem Horn, Sebastian [email protected],24 [email protected],46 Kleinberg, Jon Hotaling, Jared [email protected],44 [email protected],33 Kleinberg, Robert Houpt, Joseph [email protected],44 [email protected], 39, 50 Komarova, Natalia Howard, Marc [email protected],17 [email protected], 6, 33 Koop, Gregory Hsu, Yung-Fong [email protected],50 [email protected],48 Kosaraju, Revanth Hu, Xiangen [email protected],44 [email protected], 17, 43 Kriegeskorte, Nikolaus Huang, Nicholas [email protected],29 [email protected],9 Krishnan, Anjali Huber, David [email protected], 29, 51 [email protected],5 Kruschke, John Hurtado, Rafael [email protected], 23, 48 [email protected],42 Krusmark, Michael [email protected],35 Idrobo, Fabio Kujala, Janne [email protected],9 janne.v.kujala@jyu.fi,22 Iverson, Geoff Kwon, Taekyu [email protected],18 [email protected], 28, 29

Jameson, Kimberly A. Lake, Brenden [email protected],17 [email protected],36 Jastrzembski, Tiffany Lamport Commons, Michael tiff[email protected],35 [email protected],21 Jerger, James Lee, Michael [email protected],51 [email protected], 14, 24, 39, 46, 51 Jimison, Holly Lee, Soo-Young [email protected], 15, 23, 25, 47 [email protected],13 Johnson, Joseph Levy, Lawrence [email protected],50 [email protected],51 Johnson, Matthew Li, Yunfeng [email protected],38 [email protected],29

55 Li, Yunfeng Neufeld, Richard [email protected], 13, 28 [email protected], 34, 51 Lim, Shiau Hong Neumann, Jane [email protected],18 [email protected],19 Liu, Yang Newell, Ben [email protected],47 [email protected],15 Lo, Siu Lung Ohl, Frank [email protected],9 [email protected],47 Love, Bradley Olson, Kenneth brad [email protected],19 [email protected],52 Luce, R. Duncan Opfer, John [email protected],12 [email protected],41 Madan, Christopher Orozco-Hormaza, Mariela [email protected],46 [email protected],42 Maloney, Larry Pavel, Misha [email protected],2 [email protected], 15, 23, 25, 47 Marvel, Seth Perfors,Amy [email protected],44 [email protected],3 Maurits, Luke Perry, Lacey [email protected],3 [email protected],40 McGuire, George Peruggia, Mario [email protected],51 [email protected],38 McKanna, James Petrov, Alexander [email protected] ,25 [email protected],40 Mei, Yajun Piantadosi, Steven [email protected] ,41 [email protected],3 Merkle,Ed Pitt, Mark [email protected] , 11, 28 [email protected], 21, 41, 49 Messner, William Pizlo, Zygmunt [email protected], 16, 18 [email protected], 13, 28, 29 Miller, Brent Pleskac, Timothy [email protected], 12, 24 [email protected], 31, 42 Morgan, James Polyn, Sean james [email protected],4 [email protected],8 Muniyappa Mahendra, Koralur Pooley, James [email protected],48 [email protected],24 Myung,Jay Popov, Sergey [email protected], 21, 41, 49 [email protected],39 Popova, Anna Narens, Louis [email protected], 16, 39 [email protected], 12, 16 Prince, Melissa Navarro, Daniel [email protected],33 [email protected],3 Purdy, Brendan Negen, James brendan [email protected], 16, 17 [email protected],45 Nelson, Angela Quimbay, Carlos [email protected], 6, 7, 31 [email protected],42

56 Quinn, Max Shankle, William [email protected],47 [email protected],24 Sheu, Ching-Fan Ratcliff, Roger [email protected],28 ratcliff[email protected], 9, 10, 49 Shi, Jenny Reagor, Mary Kathryn [email protected],46 [email protected],51 Shi,Yun Regenwetter, Michel [email protected], 28, 29 [email protected], 16, 18, 39 Shiffrin, Richard Rho, Yun Jin shiff[email protected], 6, 7, 20, 26, 31, 33, [email protected],38 36 Richardson, Andrew Sieck, Winston [email protected],21 [email protected],28 Rieskamp, J¨org Simen, Patrick [email protected],27 [email protected],35 Rodgers, Stuart Smith, Linda [email protected],35 [email protected],32 Rubin, Timothy Smith, Philip [email protected],51 [email protected],14 Smith, Rebekah Salakhutdinov, Ruslan [email protected],46 [email protected],4 Smithson, Michael Sanborn, Adam [email protected],41 [email protected],30 Sreekumar, Vishnu Sanderson, Yasmine [email protected],8 [email protected],20 Steingrimsson, Ragnar Santos, Alex [email protected],12 [email protected],42 Steyvers, Mark Sarnecka, Barbara [email protected], 12, 24, 30, 36, 37, [email protected],45 39, 52 Sawada, Tadamasa Stollhoff, Rainer [email protected], 13, 28, 29 rainer.stollhoff@mis.mpg.de,37 Sch¨afer, Andreas Stringer, Sven [email protected],19 Scheibehenne, Benjamin [email protected],39 Strogatz, Steven [email protected],27 Schweickert, Richard [email protected],44 [email protected],40 Sebastian, Stephen Tang,Yun [email protected], 13, 28 [email protected],41 Serences, John Tenenbaum, Joshua [email protected],10 [email protected], 3, 4, 36 Sewell, David Thomas, Robin [email protected],14 [email protected],50 Shanahan, Matthew Tomlinson, Tracy [email protected],34 [email protected],5 Shankar, Karthik Torralba, Antonio [email protected], 6, 33 [email protected],4

57 Townsend, James Weinshall, Daphna [email protected], 39, 50 [email protected] ,47 Trueblood, Jennifer Wershbale, Avishai [email protected],32 [email protected],42 Tuerlinckx, Francis White, Corey [email protected], 8, 10 [email protected],10 Turner, Brandon White, Halbert [email protected],38 [email protected],16 Turner, Robert [email protected],19 Xi, Zhuangzhuang [email protected],40 van der Maas, Han Yao, Weiguang [email protected],44 [email protected],51 van Dongen, Hans Yechiam, Eldad [email protected],9 [email protected],23 van Ravenzwaaij, Don Young, Christopher [email protected],44 [email protected],41 Van Zandt, Trisha Yu, Chen [email protected], 38, 52 [email protected], 32, 36 Vandekerckhove, Joachim Yurovsky, Daniel [email protected],8, [email protected],32 10 Verheyen, Steven Zand Scholten, Annemarie [email protected],8 [email protected],31 Verkuilen,Jay Zeigenfuse, Matthew [email protected],41 [email protected],37 Viau-Quesnel, Charles Zhang, Shunan [email protected],40 [email protected], 39, 46 Vigliocco, Gabriella Zhuang, Yuwen [email protected],3 [email protected],8 Vigo, Ronaldo Zwilling, Chris [email protected], 14, 48 [email protected],16 Vollick, Daniel [email protected],51 von Helversen, Bettina [email protected],35 Vul,Ed [email protected],11

Wagenmakers, Eric-Jan [email protected], 10, 19, 20, 36, 44, 49 Wang, Ranxiao Frances [email protected],18 Webb, James [email protected],29 Weidemann, Christoph [email protected],15

58 We ask that the last presenter in each session act as chair, and keep time for the session.

Time Sunday, Queen Marie (A) Sunday, Gevurtz (B) Sunday, Fireside (C) 9:00 Vandekerckhove: A cognitive psy- Cavagnaro: Adaptive Design Op- Regenwetter: Quantitative Testing chometrical model for speeded se- timization for Testing Generalized of Decision Theories: I. Probabilis- mantic categorization decisions Utility Theories tic Specification 9:20 Ratcliff: Modeling the Psychomotor Tang: Adaptive Optimal Inference Zwilling: Quantitative Testing of Vigilance Test for Evaluating Sleep and Feedback for Numerical Repre- Decision Theories: II. Empirical Deprivation sentational Change Testing 9:40 Criss: Using the diffusion model to Rho: Adaptive Strategy Choice in Popova: Concensus among Concen- discriminate between memory mod- Cognitive Diagnosis: An Applica- sus Methods els tion to Fraction Subtraction 10:00 Idrobo: Application of the Ratcliff Yurovsky: Linking Learning to Petrov: Symmetry-Based Method- Diffusion Model to Reaction Times Looking: Model Selection for Eye ology for Decision-Rule Identifica- in a Two-Choice Auditory Detection Movements tion in Same-Different Experiments Task 10:20 coffee break 10:50 Dutilh: Practice decomposed: stim- Brown: Towards a unified account Luce: Subjective Loudness As A Ra- ulus specific versus task specific ef- of decisions from experience and de- tio Scale Function of Both Intensity fects scription And Frequency 11:10 Ho: Combining fMRI and the LBA Scheibehenne: How many tools to Dzhafarov: A Criterion and Tests model to ascertain a supramodal include into an adaptive toolbox? A for Selective Probabilistic Causality accumulator in perceptual decision Bayesian approach making 11:30 White: Developing and testing RT Pleskac: Judgment Field Theory: A Smithson: Probability judgments models for new paradigms computational model of subjective and partition dependence under probabilities sample space indeterminacy 11:50 Simen: Reward maximization, drift- Wershbale: Assessment Rate in Se- Narens: A new probability theory diffusion and inter-response times in quential Risky Decision Making for belief and decision instrumental conditioning 12:10 Donkin: Diffusion versus Lin- Yechiam: The effect of losses on Perry: Matching Regularity for 2D ear Ballistic Accumulation: Dif- maximization: Loss aversion versus Shapes and Locations ferent Models for Response Time, a global-effect model Same Conclusions about Psycholog- ical Mechanisms? 12:30 lunch break (jmp meeting)

13:45 Larry Maloney: Movement Planning under Risk, Decision Making under Risk

14:45 coffee break 15:20 Batchelder: Cultural Consensus Shi: Veridical binocular perception Zand Scholten: Questionnaire data Theory of metric properties of 3D shapes are linear enough: Inferences about interactions based on polytomous items 15:40 Merkle: Hierarchical models for im- Wang: Human four-dimensional Cheng: Cognitive Modeling as De- proving individuals’ subjective prob- spatial judgments on hyper-volume rived Measurement abilities 16:00 Vul: Wisdom of the Crowd Within: Kwon: Human recovery of the shape Zeigenfuse: A Three-Parameter Sampling in Human Cognition of a 3D scene Item Response Model of Matching 16:20 Greene: Representing consensus Li: A computational model of the van Ravenzwaaij: Model Equiva- and divergence in communities of recovery of a 3D scene from a single lence between the Diffusion Model Bayesian decision-makers 2D camera image and the LCA 16:40 Steyvers: A Bayesian approach to Sawada: Any 2D image is consistent aggregate knowledge: Comparing with a 3D symmetrical interpreta- groups of independent and commu- tion nicating individuals 17:00 poster session We ask that the last presenter in each session act as chair, and keep time for the session.

Time Monday, Queen Marie (A) Monday, Gevurtz (B) Monday, Fireside (C) 9:00 Huber: Learning to Forget: An Darlow: A Rational Model of Atti- Chechile: Fundamental Properties Interference Account of Cue- tude Polarization of Reverse Hazard Functions Independent Forgetting 9:20 Davelaar: A REM model of Denton: Exploring Active Learning Mei: The 2-CUSUM Process Model retrieval-induced forgetting in a Bayesian Framework for Two-Choice Reaction Time: An Information-Discounting and Memory-Forgetting Model 9:40 Criss: Using cued recall to move be- Merkle: A multivariate hierarchical Houpt: The Statistical Properties of yond task specific models Bayesian model of confidence and the Survivor Interaction Contrast accuracy in probabilistic category learning 10:00 Howard: A distributed representa- Love: Looking to Learn, Learning to Heathcote: The Lognormal Race tion of remembered time Look: Attention Emerges from Cost Model of Choice RT Sensitive Information Sampling 10:20 coffee break 10:50 Nelson: The SARKAE model: Ex- Pavel: Computer-Based Assessment Hawkins: A particle filter account of perience with novel characters deter- of Working Memory Using Com- multi-alternative decisions mines recognition memory. Pt I puter Games 11:10 Shiffrin: The SARKAE model: Ex- Jimison: Computer-Based Assess- Austerweil: Feature learning as non- perience with novel characters deter- ment of Planning and Problem Solv- parametric Bayesian inference mines recognition memory. Pt II ing 11:30 Polyn: Extending the Context Ambert: Toward a Cognitive Model Lake: Discovering Structure by Maintenance and Retrieval model of of a Computer-Based Trail Making Learning Sparse Graphs free recall Test 11:50 Dennis: What is context? McKanna: Generalized Assessment Turner: Hierarchical Approximate of Divided Attention Through Un- Bayesian Computation obtrusive Monitoring of Computer- ized Tasks 12:10 Hemmer: Estimating Individual Sanborn: Learning of dimensional Differences in a Rational Model of biases in human categorization Memory 12:30 lunch break (smp meeting)

13:45 Mike Jordan: Recent Developments in Nonparametric Bayesian Inference

14:45 coffee break 15:20 von Helversen: Averaging and Krishnan: Distance-based Partial Jameson: Evolutionary models of Choosing in the Mind: Should Least Squares Regression (DISPLS): color categorization based on realis- Judgments Be Based on Exemplars, A new approach to relate distances tic observer models and population Rules or Both? in psychology and neuroimaging heterogeneity 15:40 Anders: Rank-Aggregation and Dunlop: How to analyze multifacto- Kaznatcheev: Evolution and cogni- Consensus in Ballroom Competition rial neuroimaging experimental de- tive cost of ethnocentrism signs with pattern classifiers 16:00 Zhang: The Wisdom of Crowds in Golden: Application of a Robust Sanderson: Color charts, aesthetics, Solving Bandit Problems Differencing Variable (RDV) Tech- and subjective randomness nique to the Department of Veterans Affairs Learners’ Perceptions Survey 16:20 Miller: A Bayesian Analysis of the Stollhoff: Measuring abnormality Hurtado: A statistical mechanics Wisdom of Crowds Within in neuropsychological deficits using treatment of children’s learning pro- generalized linear mixed models cess to write Arabic numerals 16:40 Lee: Cognitive Models, the Wisdom Addis: Calculating Prediction Inter- of Crowds, and Sports Predictions vals for Future Performance 17:00 business meeting (followed by banquet) We ask that the last presenter in each session act as chair, and keep time for the session.

Time Tuesday, Queen Marie (A) Tuesday, Gevurtz (B) Tuesday, Fireside (C) 9:00 Perfors: New computational ap- DeCarlo: On Incorporating Item Ef- proaches to word learning fects into Signal Detection Theory 9:20 Andrews: Learning Word Mean- Weidemann: ROC analyses of con- Commons-Miller: Can Perceived ings from Sequential and Syntactic fidence responses and reaction times Value Be Explained by Schedules of Statistics in recognition memory Reinforcement? 9:40 Piantadosi: Beyond Boolean Logic: Pooley: Modeling Multitrial Free Gane-McCalla: Comparing Three A learning theory for complex com- Recall Different Equations Representing positional concepts Utility for a Single Reinforcement Schedule 10:00 Salakhutdinov: One-Shot Learning Prince: The Contribution of Study- Commons: Does Hierarchical Com- with a Hierarchical Nonparametric Test Lag to Old Item Variability in plexity of Items Predict Synchrony Bayesian Model Recognition Memory Across Content and Gaps Between Stages? 10:20 coffee break 10:50 Frank: Early word learning through Nelson: Modeling knowledge forma- Fields: Structure from Sequence: A communicative inference tion and frequency effects in lexical Nonparametric Bayesian Analysis of decision Human Sequence Learning 11:10 Feldman: Using a developing lexi- Cox: Developing Knowledge and Shanahan: Mathematical Expec- con to constrain phonetic category Probing Its Structure with Primed tation of Threat, Unpredictabil- acquisition Lexical Decision ity, and Mechanisms of Stochastic “Faint Threat” in a Model of Deci- sional Control 11:30 Hills: Semantic Associates in Moth- Hu: Binarizing Multi-link Multino- Wagenmakers: Cortico-Striatal erese: Comparing the Associa- mial Processing Tree (MMPT) mod- Connections Predict Control over tive Structure of Adult and Child- els Speed and Accuracy in Perceptual directed Speech with Graph Analy- Decision Making sis 11:50 Kachergis: Modeling the Acquisi- Batchelder: A String Language for Trueblood: Explaining Order Ef- tion of the Mental Lexicon Multinomial Processing Tree Models fects. The Belief-Adjustment Model Versus The Quantum Inference Model 12:10 Andrews: A Hierarchical Logistic- Purdy: Multinomial Processing Tree Iverson: Remarks on quantum cog- Wald Model for the Analysis of Models and Bayesian Networks: nition Lexical-Decision Data Some formal results 12:30 lunch break

13:45 Thomas Griffiths: Using Probabilistic Models of Cognition to Identify Human inductive Biases

14:45 coffee break 15:20 Smith: An Integrated System Anal- Caplan: The convolution operation Vigo: A non-probabilistic measure ysis of Gain and Orienting Models of for associative memory: (nearly) of relational information based on Attentional Selection found in the brain Categorical Invariance Theory 15:40 Shiffrin: Guided Visual Search Shankar: A distributed representa- Stringer: A Bayesian model of judg- tion of internal time leads to scale ment revisions in a social setting invariant timing behavior 16:00 Hotaling: Models of Information Bracis: Rats, non-rewards, and de- Davis-Stober: Changing Minds: Integration in Perceptual Decision caying uncertainties Estimation of Multiple Preference Making States Via Normalized Maximum Likelihood 16:20 Marvel: A Dynamic Model for the Schweickert: Regularities in Dream Lee: Emergence of Cooperation be- Emergence of Two-sided Social Con- Social Networks tween Intelligent Individuals with flicts Personality 16:40 17:00 end of meeting Poster Session Kosaraju: The Predictability and Negen: An ANS-Based Model Beaton: Predicting behavior from Abstractness of Language: A Study of Children’s Small-Set Size Judg- genetics with correspondence anal- in Understanding and Usage of the ments ysis English Language through Proba- bilistic Modeling and Frequency Cadez: Mathematical Modeling of Madan: Model-mechanisms for ef- Horn: What Can the Diffusion Memory Evolution - Pure Decay fects of item- on association-memory Model Tell Us About Prospective in paired-associate learning Memory? Shi: The Wisdom of the Crowd Quinn: Detection of Rare Events Liu: Instruction determines the er- Playing the Price is Right and Categorization ror pattern in relative order judg- ments Chang: Assessing psychometric Vigo: An Extension of the Categor- Muniyappa Mahendra: Applica- functions and thresholds using ical Invariance Model to Continuous tions of Shannon’s Entropy Theo- unforced-choice adaptive methods Domains rem to Psychology Myung: Cognitive Modeling Repos- Boldt: Modeling the Time Course of Dodds: Bow Effects in Absolute itory Interference Identification Houpt: Functional Principal Com- Koop: Response dynamics and the Brunton: Assessing perceptual ponent Analysis and the Capacity time course of risky decision making correlations in the uncertainty Coefficient paradigm: Exploring predictions of decision models within the multidimensional signal detection framework Neufeld: “Hot Cognition” is Not Reagor: How is Story Line Depen- Rubin: Wisdom of the Crowds Ef- Beyond the Reach of Formal Mod- dent Context Represented in the fects in Prediction of Future Out- els: Higher-Dimensional Nonlinear Brain? An investigation using EEG comes; A case study using Fantasy Bifurcations Depict Psychological and MUBBADA Football Rankings Stress and Coping Olson: Bayesian Hierarchical Model Evaluation for the Receiver Oper- ating Characteristic in Recognition Memory