Quick viewing(Text Mode)

The Perception of Causality: Feature Binding in Interacting Objects

The Perception of Causality: Feature Binding in Interacting Objects

The of : Feature binding in interacting objects

John K. Kruschke and Michael M. Fragassi Dept. of Indiana University Bloomington, IN 47405 [email protected] http://www.indiana.edu/ kruschke/home.html

Abstract just after impact (approximately 200ms), the is phenomenologically duplicitous: It belongs to the ®rst When one billiard ball strikes and launches another, most ob- while the second object has it. Thereafter, the motion of the servers report seeing the ®rst ball cause the second ball to move. Michotte (1963) argued that the essenceof phenomenal second object becomes autonomous. causality is ªampliationº of movement, in which the motion of Unfortunately, Michotte made only phenomenological ob- the ®rst object is perceptually transferred to the second object. servations of perceived causality and gathered no perfor- Michotte provided only phenomenologicalevidence,however. mance data, leaving open the possibility that ampliation of We extend the reviewing of Kahneman, Treisman, andGibbs (1992)to Michotte- launchingeventsand report motion is merely an idiosyncratic epiphenomenon (Boyle, response-time data consistent with Michotte's of ampli- 1972; Joynson, 1971). Nevertheless, Michotte's methods and ation. We discuss how contemporary theories of feature bind- ®ndings are frequently described in textbooks on perception, ing can extend to the domain of interacting objects and address development, arti®cial , etc. (e.g., Boden, 1977; our results. We also suggest that our treatment of ampliation Bower, 1982; Bruce & , 1990; Rock, 1975), and a com- helpsclarify controversiesregarding whetherperceivedcausal- ity is direct or interpreted and whether it is innate or learned. pilation of newly translated articles by Michotte has recently been published (ThinÁes, Costall, & Butterworth, 1991). In this paper we report a response-time experiment that yielded Causality Perceived results consistent with Michotte's theory of ampliation. Imagine a billiard table. The cue ball rolls across the felt and We propose that the theory of ampliation can be construed strikes the 8-ball, launching the 8-ball. Most observers will to impact directly on theories of feature integration; i.e., the- report seeing the cue ball cause the 8-ball to move. The per- ories of how different visual features of an object, such as ception of causality has been placed at the foundationof cog- shape, , and movement, are bound into a common iden- nition: Kant (1788) argued that causality was an innate and tity but distinguished from visual features of other objects in fundamental category of ; Piaget (1971) made it an the same ®eld of view (e.g., Treisman, 1986). The key is integral part of his theory of development; and, more recently thatifthemotionis perceptuallytransferred fromoneobjectto Leyton (1992) has argued that the extraction of causality is at the next, then the feature of movement must be unboundfrom the core of perception and cognition. the launching object and bound with the launched object. We The British empiricist philosopher (1739) ar- discuss how contemporary theories of feature binding can ac- gued that impressions of causality are mere fabrications of a count for Michotte's phenomenology of perceived causality. sophisticated : In the case of the billiard balls, the ob- Providing a performance measure of ampliation and giving server sees the cue ball move, sees the contact of the balls, it a theoretical interpretation in terms of feature binding also and sees the subsequent motion of the 8-ball,but does not see suppliesa new perspective on the relation between ampliation causality itself. The impression of causality, Hume argued, and perceived causality. Rather than debate whether a single is a learned nexus from the ®rst to second ball, based on re- process of perceiving causality is either innate or learned, we peated of the conjunction of the two , suggest that the sub-process of ampliation might be perceived their spatio-temporal contiguity, and the temporal priority of directly and developed early in infancy, but the complete per- the one motion relative to the other. ception of causality might be interpreted and learned. Two hundred years later, the Belgian Albert Michotte impugned Hume, arguing instead that the impres- An Empirical Approach to Measuring sion of causality is a spontaneous perceptual gestalt, which is Ampliation neither learned nor an interpretation via abstract of physical events (Michotte, 1941, 1963). Michotte claimed The Reviewing Paradigm that the of perceived causality is ªampliation of mo- The performance measure of ampliation that we will describe tion.º The neologism, ªampliation,º refers to two aspects of is an extension of the reviewing paradigm invented by Kah- theperceived motion. First, themotionof theapproachingob- neman et al. (1992). In this paradigm, the observer is ®rst ject is transferred to the launched object. Second, for a brief shown two objects on a computer screen, such as the triangle P1

T

P2

Preview Field Linking Display Target Field

Figure 1: Example of the reviewing paradigm used by Kahneman et al. (1992).

and square in the left panel of Figure 1. Two letters, labeled In an event we call launching, shown in the left panel of Fig- ªP1º and ªP2º in Figure 1, are brie¯y presented (ªpreviewedº) ure 2, participants saw one circle move toward and contact inside each object. After the letters disappear, the empty ob- another, at which time the ®rst circle stopped and the sec- jects visibly move to new locations, as shown in the middle ondcircle moved away at the same velocitypreviouslyhad by panel of Figure 1. The motion of the objects linkstheir initial the ®rst circle. Analogous to the reviewing paradigm of Kah- positions to their ®nal positions, and so this motion is called neman et al. (1992), symbols such as ª@º or ª&º appeared the ªlinking display.º A target letter, labeled ªTº in the right brie¯yinthe initialmoments of the event, indicated byP1 and panel of Figure 1, then appears in one of the objects. The ob- P2 in Figure2. Then oneempty circle launched the other, and server's task is to identify the letter as quickly as possible, by a target symbol, indicated by T in Figure 2, appeared in the saying the letter's name into a microphone. launched object, at which time the launched object was sta- The key result is that observers can identify target letters tionary. Unlike the identi®cation task used in the experiments that matched the preview letter from the same object faster of Kahneman et al. (1992), the task for our participants was than they can identify target letters that matched the preview to indicate as quickly as possible whether the target symbol letter from the other object. In Figure 1, for example, ob- was the same as either of the two preview symbols. Responses servers are faster to identify the target, T, when it matches were made by pressing a button when the target matched ei- P2 (which was in the same object as the target) than when therof thepreviewsymbols, and by pressing adifferent button it matches P1 (which was in the other object). Kahneman when the target did not match either of the preview symbols. et al. (1992) called this effect the object-speci®c preview ad- A second event type, called delayed motion, began and vantage. ended the same way as launching, but had the two circles remain in contact with each other for approximately 890ms. The Reviewing Paradigm Applied to Launching According to Michotte (1963, Experiment 29, p. 91), ob- Kruschke(1987)applied thereviewingparadigm to Michotte- servers perceive delayed motion as two independent move- style launching events. Suppose the linking display in the re- ments withoutampliation. The ®rst circle is seen to stop com- viewing paradigm did not keep the objects separated, as in pletely, and the second circle then appears to move away with Figure 1, but instead showed one object striking and launch- its own motion. Because the two motions are perceived as in- ing the other, as one billiard ball can strike and launch an- dependent, we would expect to ®nd a robust object-speci®c other. Consider what would happen if the target letter ap- preview advantage in delayed motion. In launching,however, peared in the launched object. Would there still be a strong we predicted that the object-speci®c preview advantage would object-speci®c preview advantage, or would the preview in- be diminished. formation from the launching object be transferred, or am- Participants also saw two other events in which the im- pliated, to the launched object? Kruschke (1987) reported pacted objectdid notmove. The thirdpanel of Figure2 shows that the object-speci®c preview advantage was signi®cantly the event we call target at contact, and the fourthpanel shows reduced in launching, relative to a control event in which the delayed target at contact. In these events, the target appeared objects did not interact. at the moment corresponding to when the ®rst circle contacted We replicated and extended that study in new experiments. the second circle in the launching event. The for Time P1 P2 P1 P2 P1 P2 P1 P2

T T

T T

T T T T

Launching Delayed Motion Target at Contact Delayed Target at Contact

Figure 2: Schematic diagram of the four events in our experiment (not drawn to scale).

these events was to encourage observers to attend to the point tion, centered laterally on the screen with a small yellow ®xa- of impact (which is where the target appears in these ªcontactº tion dot centered between them. One circle and symbol were events). Previous experiments suggested that if observers see red, the other green. The ®xation dot and preview symbols only the launching and delay events, they might immediately appeared for 500ms. The preview symbols then disappeared, move their after the preview ®eld to the anticipated and the empty circles remained stationary for 200ms. For location of the target, without devoting much attention to the launching (see Figure 2), the circles remained stationary for motions of the objects. Attending to the point of impact was an additional 890ms and then underwent the launching mo- deemed important by Michotte for eliciting the best phenom- tion for 230ms. When moving, the speed of the circles was enal ampliation; attending to the point of impact is also im- approximately 33cm/s. Then the target letter appeared in the portant to clearly distinguish launching from delayed motion. launched object. When the target letter matched a preview let- We anticipated a strong object-speci®c preview advantage for ter, the target letter had the same color as the preview letter it delayed target at contact, but a diminished advantage in target matched. The other events had the same total duration, with at contact. onlythelinkingdisplay differingbetween them. For example, in delayed motion, the linking display consisted of approxi- Method mately 115ms for the motion of the ®rst object, followed by 890ms of contact, following by 115ms for the motion of the Participants. Forty-two undergraduates at Indiana Univer- second object. sity volunteered in partial ful®llment of a psychology course requirement. Design and procedure. The design consisted of ®ve Stimuli. Stimuli were presented on a PC-type 13º color crossed factors: (1) type of match between target and pre- monitor in VGA resolution. On every trial, there was a yel- view letters (match same object, match other object, match low rectangular frame, 256mm wide by 60mm high, which neither object); (2) direction of motion (left, right); (3) color enclosed the relevant region of the the display. The circles of the left preview symbol and circle (red, green); (4) time of had a diameter of approximately 27mm, and were separated initial movement (late as in launching and target at contact, center-to-center by approximately 38mm. Every trial began early as in the delay events); and (5) time of target appearance with the circles and their preview symbols in the same posi- (late as in launching, early as in contact). The trials in which 750

700

650

Response Time (msec) 600 Match Neither Match Other Match Same

550 Launch Delay Contact Delayed Contact Event

Figure 3: Mean response for correct responses.

the target did not match a preview letter were doubled in thoughthe reductionwas statisticallyof marginal signi®cance order to equalize the total number of match and no-match (interaction contrast F(1,41) = 3.78, p = .059). Comparable trials. Each block consisted of 66 trials: 2 warm-up trials reductions have been observed in several other experiments chosen at random from the design, followed by 64 trials that conducted in our laboratory, and by Kruschke (1987), so we visited each cell of the design once. consider this trend toward reduction to be reliable. The re- After on-screen instructions with several examples of the duction in object-speci®c preview advantage was quitestrong events, participants had an initial practice block of 35 trials, for target at contact, however (interaction contrast F(1,41) = followedby8blocksof66trialseach, withbriefrestsbetween 13.50, p = .0007). Thus, at the time of impact, there was a blocks. The experiment lasted about an hour. complete loss of object-speci®c preview advantage, but mo- ments later, in the delay events, the object-speci®c advantage Results was regained. Figure3 shows the mean response times as a functionof event These results are intriguing for two reasons. First, regard- and match condition. These factors did not interact with di- less of their theoretical interpretation,the results show that the rection of motion or color of symbol. There was indeed a object-speci®c preview advantage is in¯uenced by the type of strong object-speci®c preview advantage for delayed motion interaction between the objects. In particular, mere contact and for delayed target at contact, with response times signif- of the objects, as in delayed motion, does not obliterate the icantly longer for matching the target to the other object than object-speci®c preview advantage, but contact with launch- for matching thetarget to thesame object (for delayed motion, ing does diminish the advantage. Moreover, the reduction in RTs for match-other and match-same were 703ms vs. 646ms, object-speci®c preview advantage is temporally localized just respectively, F(1,41) = 53.71, p = .0001). The magnitude of after the time of impact, and the object-speci®c advantage is the object-speci®c advantage was reduced for launching, al- regained moments later. Second, one interpretation of the re- sultsis thatinformationfromthelaunchingobjectis ampliated pulse trains. At impact, the directional of motion de- to the launched object, so that just after impact the previewed tectors causes the new movement of the launched object to symbol from the launching object is as accessible to retrieval have the same synchronization as the launching motion, so as the previewed symbol from the launched object. These re- that justafter impact the motionof thelaunched object is syn- sults are the ®rst performance measure directly addressing Mi- chronized with the launching object, but localized with the chotte's (1963) notion of ampliation in perceived causality. launched object. This accounts for the duplicity of motion in Michotte's phenomenology: The motion of the launched Ampliation as Feature Re-Binding object is localized with the launched object, but still belongs To explain their results, Kahneman et al. (1992) suggested a to (is synchronized with) the launching object. This also ac- ªreviewing model,º in which the motion of the linking dis- counts for the dif®culty and equalization of accessing the his- play (middle panel of Figure 1) which object will be tory of both objects: The target that appears in the launched reviewed ®rst in . The model assumes that the motion object at the moment of impact is not clearly synchronized links the target object with the corresponding preview object, with either pulse train, and so both preview symbols are re- but the model does not incorporate any aspects of the dynam- trieved with equal dif®culty. After a brief time, the motion ics of the motion. In particular, the model does not suggest of the launched object becomes synchronized with the other why different events would have different biasing effects with features of the launched object, and hence the object-speci®c different durations, as found in our data. Treisman and Kah- preview advantage is resurrected. neman (Treisman, 1986; Kahneman et al., 1992) have also This explanation is not committed to pulse train synchro- suggested that visual features are bound together into ªobject nization as the only possible binding mechanism. The expla- ®lesº that retain each object's immediate history. An object nation merely requires that the process of bindingtakes some ®le is tantamount to a tag, or label, on each visual feature, small but non-zero time, and that the object tag of the launch- identifying the object to which the feature belongs. ing object primes the launched object at the time of impact. Weproposethatour resultsare mediated by bindingand un- Pulsetrain synchronizationis just onepossiblemechanism for binding of the visual feature of motion with other features of implementing these . the objects. One potential (and controversial) mechanism for feature integration is synchronization of neural pulses from different feature detectors. The pulsetrain acts as an object la- Ampliation versus Causality bel for the feature, and synchronization gives the labels from different features a common signature. There is empirical ev- Michotte (1941, 1963) argued that the perception of causal- idence that neurons use this mechanism (for a recent review ity is not an interpretation based on acquired knowledge of see Singer & Gray, 1995), and several researchers have ap- mechanical events, but instead is perceived directly, with am- plied this binding mechanism in models of , pliation as its essence. Our results do not necessarily support attention, and memory (e.g. Damasio, 1990; Grossberg & thisperspective, in full. It is possiblethat ampliation, qua fea- Sommers, 1991; Hummel & Biederman, 1992; Lumer, 1992; ture unbinding and rebinding, is a direct perceptual mecha- Mozer, Zemel, & Behrmann, 1992; Pabst, Reitboeck, & Eck- nism, but the perception of causality is an additional interpre- horn, 1989; Sporns, Tononi, & Edelman, 1991). tive process. For example, Weir (1978) described a model that Applied to the scenario of launching, these suggest classi®es Michotte-style collision events without ever men- that as thelaunchingobject approaches theto-be-launched ob- tioning ampliation, presumably because of the 's in- ject, the pulse trains of the motion detectors for the motion tangibility in Michotte's theories. Rei®cation of ampliation, of the launching object are synchronized with the pulse trains as suggested by our experiments, calls out for theories to ad- of the other features of the launching object, all of which are dress it, and provides one avenue for distillingwhich aspects desynchronized from the pulse trains of the features of the to- of perceived causality are direct and which are interpreted. be-launched object. The question then becomes, What hap- Michotte also argued that the perception of causality is in- pens to the synchronization of the pulse trains when the ob- nate. Leslie (1982; Leslie & Keeble, 1987) providedevidence jects come into contact? that six-month old infants can distinguish causal from non- One answer to this question is suggested by the causal events, or at least are sensitive to reversals of agency that motion primes other motion detectors along the forward in causal events. Whether the infants perceive such events as trajectory. Long-range directional priming of motion detec- causal or not remains an open question. It might be that am- tors has been discussed extensively in models of visual motion pliation develops rapidly, like stereopsis, in response to the perception developed by Marshall (1990; Martin & Marshall, visual world, whereas interpretations of causality are learned 1993), and of long-range directional con- later in childhood. Perhaps ampliation is used as a perceptual nections between motion-detectingneurons comes from work cue for subsequent causal interpretation, and so the two are by Gabbott, Martin, and Whitteridge (1987) and others cited correlated. Separating ampliation from causality also allows by Marshall (1990). for the possibility of sensitivity to ampliation, but Our results might then be explained as follows: In the individualdifferences in the perception of causality (Beasley, launchingevent, thetwo objectsinitiallyhave desynchronized 1968; Schlottman & Anderson, 1993). Acknowledgments Leslie, A., & Keeble, S. (1987).Do six-month-oldinfantsper- This was supported in part by NIMH FIRST Award ceive causality?. Cognition, 25, 265±288. 1-R29-MH51572-01 to Kruschke, and by an Indiana Univer- Leyton, M. (1992). Symmetry, Causality, Mind. MIT Press, sity Cognitive Program Summer Research Fellow- Cambridge, MA. ship to Fragassi. Thanks to Michael Erickson, Mark Johansen Lumer, E. D. (1992).Selective attention to perceptual groups: and Nathaniel Blair for comments on a previous draft, and to The phase tracking mechanism. International Journal of Colin Bogan, Amanda Reed, and Eddy Riou for administer- Neural Systems, 3, 1±17. ing experiments. Thanks also to and for encouragement and access to their lab at U.C. Marshall, J. A. (1990). Self-organizing neural networks for Berkeley in 1986-87. perception of visual motion. Neural Networks, 3, 45±74. Martin, K. E., & Marshall, J. A. (1993). Unsmearing vi- References sual motion: Development of long-range horizontal intrin- Beasley, N. A. (1968). The extent of individual differences in sic connections. In Hanson, S. J., Cowan, J. D., & Giles, the perception of causality. Canadian Journal of Psychol- C. L. (Eds.), Advances in Neural Processing ogy, 22, 399±407. Systems 5. Morgan Kaufman. Boden, M. A. (1977). Arti®cial intelligence and natural man. Michotte, A. (1941). La causalitÂe physique est-elle une Basic Books, New York. donnÂee phÂenomÂenale?. Tijdschrift voor Philosophie, 3, Bower, T. G. R. (1982). Development in infancy. W. H. Free- 290±328. Unpublished translation by Krista Hensley, April man, San Francisco. 30, 1993. Boyle, D. G. (1972). Michotte's ideas. Bulletin of the British Michotte, A. (1963). The perception of causality. New York: Psychological Society, 25, 89±91. Basic Books. Translated from the French by T. R. and E. Miles. Bruce, V., & Green, P. R. (1990). Visual perception: physiol- ogy, psychology & ecology. Erlbaum, East Sussex, UK. Mozer, M. C., Zemel, R. S., & Behrmann, M. (1992). Learn- ing to segment images using dynamic feature binding. Neu- Damasio, A. R. (1990). Synchronous activation in multiple ral Computation, 4, 650±665. cortical regions: a mechanism for . Seminars in the , 2, 287±296. Pabst, M., Reitboeck, H. J., & Eckhorn, R. (1989).Amodel of preattentive region de®nition based on texture analysis. In Gabbott, P. L. A., Martin, K. A. C., & Whitteridge,D. (1987). Cotterill,R.M. J.(Ed.), Modelsof BrainFunction,pp.137± Connections between pyramidal neurons in layer 5 of cat 150. Cambridge University Press, Cambridge, England. (area 17). Journal of Comparative Neurology, 259, 364±381. Piaget, J. (1971). Les Explications Causales. Presses Univer- sitaires de France, Paris. Translation by D. & M. Miles, Un- Grossberg, S., & Sommers, D. (1991). Synchronized oscilla- derstanding Causality. New York: Norton. 1974. tions during cooperative feature linking in a cortical model of visual perception. Neural Networks, 4, 453±466. Rock, I. (1975). An introduction to perception. Macmillan, New York. Hume, D. (1739). A treatise of . Oxford, Eng- land: Oxford University Press. Republished 1980. Schlottman, A., & Anderson, N. H. (1993). An information integration approach to phenomenal causality. Memory & Hummel, J. E., & Biederman, I. (1992). Dynamic binding in Cognition, 21, 785±801. a neural network for shape recognition. Psychological Re- view, 99, 480±517. Singer, W., & Gray, C. M. (1995). Visual feature integration and the temporal correlation hypothesis. Annual Review of Joynson, R. B. (1971). Michotte's experimental methods. , 18, 555±586. British Journal of Psychology, 62, 293±302. Sporns, O., Tononi, G., & Edelman, G. M. (1991). Model- Kahneman, D., Treisman, A., & Gibbs, B. J. (1992). The re- ing perceptual grouping and ®gure-ground segregation by viewingof object ®les: Object-speci®c integrationof infor- means of active reentrant connections. Proceedings of the mation. , 24, 175±219. National Academy of USA, 88, 129±133. Kant, I. (1788). . St. Martin's Press, ThinÁes, G., Costall, A., & Butterworth, G. (Eds.). (1991). Mi- New York. Translated by N. K. Smith, 1965. chotte's experimental phenomenology of perception. Erl- Kruschke, J. K. (1987). The perception of causality: A per- baum, Hillsdale, NJ. formance measure of ampliation.. Paper presented at the Treisman, A. M. (1986). Features and objects in visual pro- Ninth Annual Berkeley-Stanford Conference in Cognitive cessing. Scienti®c American, 255(5), 114B±125. Psychology. Available from the author. Weir, S. (1978).Theperception of motion: Michotterevisited. Leslie, A. (1982). The perception of causality in infants. Per- Perception, 7, 247±260. ception, 11, 173±186.

Top View