Intelligent Systems Conference 2018 6-7 September 2018 | London, UK A Computer Science Perspective on Models of the Mind

Teresa Nicole Brooks∗, Abu Kamruzzaman†, Avery Leider‡ and Charles C. Tappert§ Seidenberg School of Computer Science and Information Systems, Pace University Pleasantville NY Email: ∗[email protected], †[email protected], ‡[email protected], §[email protected]

Abstract—One area of intense focus in Artificial Intelligence Our hypothesis is that computer science based theories for (AI) research is to implement intelligent agents and machines creating human-like minds are likely better suited for modeling that can think, reason, and solve problems with similar if not learning related functions, such as object and abstract concept better proficiency of human beings. The advancements in our recognition as well as modeling memory for storing low-level understanding of intelligence and its governing principles have to high-level abstract concepts. These theories are strongly lead researchers to explore vastly, different and passionately based on our current understanding of the human mind. In debated approaches to building intelligent systems. However, in recent years a consensus appears to be emerging as observed by particular, the structure and functions of the neocortex as the 2013 AAAI Fall Symposium on Integrated Cognition, where they encapsulate properties such as variation-tolerance, self- an initial proposal is for a standard model of human-like minds. organization, and the ability to continuously adapt to new This model is largely based on current state-of-the-art cognitive input. They also have proven to be practical by the success architectures, such as ACT-R, Sigma and Soar. While we do not of closely related algorithms such deep neural networks and disagree with the proposed model, we believe more computer Hidden Markov Models which have been used to for learning science based theories for modeling human-like minds should be tasks such as pattern and . considered for modeling memory and recognition functions. In this paper, we present a survey of the current literature that The goal of this paper is not to offer detailed changes to the supports our position. currently proposed standard model of the mind, but to present a Keywords—Standard model of the mind; cognitive architectures; survey of the current literature that supports our position that artificial intelligence; artificial neocortex; deep learning computer science based approaches to modeling human-like minds should be considered for modeling memory and some recognition functions. I.INTRODUCTION The remainder of this paper is organized as follows. Section Artificial Intelligence (AI) research is unique as it often II provides a summary of related work. In Section III, what requires cross-discipline knowledge from fields such as linguis- we currently know about the human mind is briefly discussed. tics, cognitive science, and neuroscience. As our understanding Section IV highlights key aspects of leading, proposed com- of what constitutes intelligence and the fundamental principles puter science based theories for developing artificial minds. In that governs its growth, researchers have been trying to de- Section V, we discuss the relationships between deep learning velop systems that can reason, learn and solve problems like and architecture. The last sections contain proposed humans. This desire has fueled decades of contentious debates, future work and conclusions. regarding the vastly different approaches to build intelligence systems, but in recent years there appears to be an emerging consensus regarding the current state of the field among the AI research community [1]. This consensus was observed during II.RELATED WORK the 2013 AAAI Fall Symposium on Integrated Cognition and In this section, we briefly discuss related research focused ultimately led to an initial proposal for a standard model of on the common threads between computer science, neuro- human-like minds. Since this symposium, there has been a science, and computing. call to engage the international research community to further develop a standard model of the mind. The initially proposed standard model and the extended A. Human Brain Project proposal by Laird et al. is based on the hypothesis that cognitive architectures, such as ACT-R [2], Sigma [3], and In 2011, Markam et al. introduced the Human Brain Project Soar [4] provide abstractions for functions such as memory, [6]. This project’s primary objective is to drive innovation in learning and perception [1]. While we do not disagree with the fields of computing, brain-related medical research, and the currently proposed model, we do propose that the research neuroscience. The project seeks to integrate research from community explore computer science based theories for mod- various disciplines under a single platform to learn more about eling human-like minds. In particular, theories that propose the functions and structure of the brain, with the ultimate models for implementing an artificial neocortex, the part of the goal of using these learning’s to build models that can be mammalian brain that is considered to be where all “thinking” tested and validated by running them against simulations on takes place [5]. supercomputers. 978-1-5386-4185-9/18/$31.00 c 2018 IEEE 1 | P a g e Intelligent Systems Conference 2018 6-7 September 2018 | London, UK

B. Biologically Inspired Models We understand the mind to be “the faculty of and thought” [9]. The mind is capable of perception, judg- Fontana 2017, introduced a storage device model inspired ment, imagination, recognition, memory, feelings, attitudes, by the hippocampus. During this research, similarities were and emotions. All of these things determine our choice of observed between deep neural networks and biological neural action. Various structures of the brain are responsible for networks of the hippocampus and cortex [7]. These obser- specific mental processes. For example, our consciousness is vations led Fontana to the conclusion that the delusions of affected by the prefrontal cortex, using collections of neurons schizophrenia and psychosis, as well as the depersonalized with parallel connections across other regions in the brain thoughts of PTSD, are examples of the brain coming to a [10] Opris [11] suggests that interactions between interlam- flawed decision without training on enough data, just as deep inar prefrontal microcircuits, the posterior parietal cortex, artificial neural networks produced weak models without an and cortico-striatal-thalamo-cortical circuits are responsible for adequate amount of training data. making our decisions. The neocortex of the brain is responsible for sensory perception, recognition of everything from visual C. Neoroscience and objects to abstract concepts, controlling movement, reasoning from spatial orientation, rational thought and language in what In recent years because of the popularity and success of we regard as “thinking” [5]. The neocortex may also recognize deep learning, some researchers are striving to integrate deep patterns. According to Kurzweil, this is possible because the learning and neuroscience. Because of deep learning’s origins neocortex has a columnar organization, as first discovered by in neuroscience, this is a natural progression. Researchers Vernon Mountcastle. such as Marblestone et al. argue that the advances made in machine learning research specifically in the field of artificial We can use the brain’s architecture as a blueprint for neural networks make the fields of neuroscience and machine designing a digital counterpart. Kurzweil [5] estimates the learning “ripe for convergence [8].” They also state that the mind contains “30-100 million bytes of compressed code,” synthesis of machine learning and neuroscience highlights the and artificial intelligence, if created based on this design and many differences in research language, terms, and divergent using hidden Markov models and genetic algorithms, could investment between the two disciplines, and seeks to bring surpass the human mind in its capabilities. But an artificial them back into convergence by documenting their similarities brain of that ability will require massive computational power around three hypotheses discussed in the sections below. that will not be reached for another decade [12]. The [13] has thus far only managed to replicate a 1) Brain optimizes cost functions: The first hypothesis rodent brain [6]. Fig. 1 below shows the timeline of the actual is that the brain optimizes cost functions. The term “cost and projected progress of the Blue Brain simulation project functions” is used by Marblestone et al. [8] to describe what by year in a 45-degree line on a graph. The progression in neuroscience calls “learning”. They present evidence that this the X-axis represents computer speed (FLOPS) and in the Y- cost optimizing capability, which is a powerful part of deep axis represents computer memory in bytes needed to run the learning, is also present in the brain at both the micro level of project. individual neurons and the macro level of the cortex. 2) Cost functions in the brain can be localized: The next hypothesis is that cost functions in the brain dynamically change over time and are not necessarily global functions. Marblestone et al. argue that it is possible that different cost functions optimize different things in different parts of the brain and that cost functions may change over time (i.e. as we age) [8]. 3) Functions of the brain have specialized systems: The final hypothesis is that key functions of the brain have special- ized systems and more specifically this specialization leads to efficient solutions to computational problems. Marblestone et al. argue that the difference in the flow of information in the brain could indicate that there are specialized algorithms that are used to efficiently solve specific problems in the brain.

III.WHAT WE CURRENTLY KNOWABOUTTHE HUMAN MIND Fig. 1. Actual and projected progress of the Blue Brain brain simulation Only the human brain has intelligence, meaning the ability project (adapted) [5]. to evaluate itself and change [5]. The mind is the core functional object of any mammal. There is no actual visibility Dharmendra Modha and others simulated a digital brain of mind. We understand the brain as the visible entity of the model, cell-by-cell, of a partial human visual neocortex that mind. Our understanding of the mind has improved through contains 1.6 billion virtual neurons and 9 trillion synapses studying artificial intelligence, cognitive science, neuroscience, equivalent to a cat neocortex [5]. Hierarchical Hidden Markov and robotics [1]. Models(HHMMs) [14] are used for speech recognition and 978-1-5386-4185-9/18/$31.00 c 2018 IEEE 2 | P a g e Intelligent Systems Conference 2018 6-7 September 2018 | London, UK natural language texts.

IV. THEORIESFOR DEVELOPING ARTIFICIAL MINDS The neocortex is the part of the brain where researchers believe thinking occurs. It is this belief along with discoveries regarding the structure and functionality of the neocortex that has motivated researchers to develop biologically inspired models of artificial minds (Table I). In this section, we will briefly describe key attributes of two theories for modeling artificial minds.

A. Theory of Mind (PRTM) ’s Pattern Recognition Theory of Mind is a theory for describing the basic algorithms of the neocortex. It is based on the hypothesis that the neocortex is a homogeneous, Fig. 2. Diagram of a single neocortical pattern recognition module (adapted) recursive structure that is composed of a large number of basic [5]. structural units called pattern recognizers [5], hence making the neocortex itself a pattern recognizer. This bottom-up, hierarchical organization of patterns and pattern recognizers exceeds a learned threshold, it will signal a neighboring pattern is a key attribute of this theory, as it allows for the expression, recognizer one level higher in the hierarchy that it successfully matching, and storage of complex and abstract concepts. identified it’s input pattern, which means the recognizer above Pattern recognizers are defined by a hierarchy of self- it can now process its input pattern. Parameters such as the organizing connections that link together. When a new pattern matching threshold, size and size variability are learned by is learned, new connections are formed between the pattern running genetic algorithms [5]. Pattern recognizers have the recognizers that were involved in recognizing the given input ability to send signals to other pattern recognizers below and pattern. Each pattern recognizer is responsible for identifying above them in the hierarchy. This enables the higher-level a single input pattern. There is also redundancy among pattern pattern recognizers to signal lower-level recognizers to lower recognizers to identify the same input pattern. This redundancy their matching thresholds because most of a given input pattern allows for generalization of pattern identification, which allows has been identified. a system to learn and tolerate variations in input patterns. The self-organization of pattern recognizers is an important feature of PRTM, as it enables a system to create new connections and remove obsolete connections while it learns new input patterns over time. Kurzweil proposes using hierarchical hidden Markov mod- els (HHMM) to implement “self-organizing hierarchical pat- tern recognition” [5]. Hierarchical hidden Markov models is a statically model where each state is its own self-contained probabilistic model and each state yields a sequence of ob- servations symbols instead of an individual symbol [16]. In PRTM based systems hierarchical hidden Markov models each internal state on each level represents a single neocortical pattern recognition module as depicted in Fig. 2. States on each level can identify redundant but similar patterns. In the example shown in Fig. 3, Level 1 represents a low- level input pattern. Levels 2 and 3 represent higher-level input patterns or concepts that contains input patterns from one level below it. Each white circle labeled PRi,j represents a pattern recognition module. The gray lines represent vertical state transitions and the black lines represent horizontal state transitions, where which transition has a calculated probability.

When a pattern recognizer receives an input pattern it Fig. 3. Example of a hierarchical hidden Markov model implementing a calculates a score (probability). This score is calculated by hierarchy of redundant pattern recognition modules in a PRTM system. matching observed magnitudes of each feature of an input pattern against learned size and size variability parameters Input patterns are one-dimensional vectors and like the that are associated with this features. This score is used to pattern recognizers, they are organized in a hierarchy of determine if a pattern recognizer was successful at identifying lower and higher level patterns. Multidimensional input pattern the single input pattern it is assigned to recognize. If the score features are reduced to one-dimension vectors by a sparse 978-1-5386-4185-9/18/$31.00 c 2018 IEEE 3 | P a g e Intelligent Systems Conference 2018 6-7 September 2018 | London, UK

TABLE I. SUMMARY OF BIOLOGICALLY INSPIRED THEORIESFOR MODELINGAN ARTIFICIAL NEOCORTEX Model Implementation Key Components Pattern Hierarchical proposes a model of an artificial neocortex that is a self-organizing, recursive, pattern matching structure where Recognition Hidden Markov the relationships between the pattern matchers are defined by their self-organizing connections [5]. In addition, Theory of The Models patterns are also modeled as recursive structures where a hierarchy of patterns enables the construction of Mind (PRTM) (HHMM) complex patterns and abstract concepts. Hierarchical Sparse proposes a “biologically constrained” [15] theory that describes the fundamental principles of the neocortex. Temporal Distributed Under this theory, the neocortex is defined as a hierarchical memory system, where memory (sensory patterns) Memory (HTM) Representations are modeled as “time changing (temporal) patterns” [15]. HTM is the core technology for building intelligent (SDRs) machines whose functionality is constrained by the fundamental principles of the neocortex. coding technique called vector quantization, which Kurzweil first used in speech recognition software [5]. This process assigns a vector to be a point in a cluster and each cluster is labeled using an integer [5]. For example, when a new input pattern arrives in a system it is assigned to a cluster and can then be identified by the integer that represents its cluster. Vector quantization is an optimization that reduces data complexity while retaining the key features that are important for recognizing a pattern.

B. Hierarchical Temporal Memory (HTM) Hierarchical Temporal Memory (HTM) is a “biologically constrained” theoretical framework that describes the funda- Fig. 4. HTM’s hierarchy of memory nodes (adapted) [15]. mental principles of the neocortex [15]. It is the successor of George and Hawkins’ memory-prediction theory [17] which defined a theoretical model of the human neocortex. Like connections (activation) of neurons is what allows the mind to Kurzweil’s Pattern Recognition Theory of Mind, it defines a efficiently learn new sequences as well as make predictions. the neocortex as a recursive, homogeneous structure which a It is these “moment-to-moment” thoughts and perceptions that hierarchical organization. define which neurons are active at any given point in time. In HTM systems information is represented as Sparse Distributed Hierarchical Temporal Memory is based on three key Representations (SDRs). SDRs are vectors of thousands of principles: common algorithms of cortical regions, hierarchy, bits, where active neurons are represented by 1s and ‘off’ and sparsely connected neurons [15]. neurons are represented by 0s, where only a small percentage For the first principle, Hawkins, J. et al., propose that of neurons are actually active. This representation not only because of the neocortex’s structural uniformity and the fact allows the system to represent sparsely connected neurons but that we now know that regions of the neocortex perform very acts as a means of data encoding as each bit encodes not similar actions there must be some fundamental algorithms only data but contextual information as well. This form of that can generate behaviors for all sensory perceptions such as representation also offers interesting mathematical properties hearing, vision, and language [15]. They also define all sensory such as it’s union property which allows systems to perform actions as “temporal inference problems” whereby patterns are efficient pattern matching by only comparing a small number a hierarchy of lower and higher level time changing patterns. of features [15]. For example, when comparing SDRs a system can easily determine how semantically different or similar they The second principle is arguably the most important prin- are. Such differences and similarities are defined by the number ciple of HTM, as it asserts that cortical regions are defined of active bits they share in the same position of the vector. by a logical hierarchy of connections whereby higher-level perceptions are derived from lower-level sensory patterns. As 1) Common approaches for developing artificial minds: depicted in Fig. 4, these sensory input patterns are processed Biologically inspired theories for developing artificial minds and passed up the hierarchy as beliefs, which are then used share common approaches. Many employ methods for sparsely by the highest levels to make predictions. Like Kurzweil’s coding input patterns in order to reduce data complexity, PRTM theory, this hierarchical organization of cortical regions while retaining key features of the pattern and making pattern and patterns allows for the expression, matching, and storage matching and storage more efficient. The neocortex, as well of complex and abstract concepts. In HTM based systems as input patterns, are often organized as a hierarchy of lower memory is organized as a hierarchy of lower and higher levels and higher level structures. This hierarchical organization is of memory, where lower-level sensory patterns are taken in as key allows the learning and representation of complex and input. These input patterns are processed and passed up the abstract concepts. Lastly, these models share common origins hierarchy as beliefs, which are then used by the highest levels with early models in machine learning such as deep learning, to make predictions. These predictions are then passed down such as Rosenblatt’s (1962) topological model of the nervous and can be used for example to trigger motor functions. system depicted in Fig. 5. Like their predecessors, modern theories for developing artificial minds build complex networks The third principle defines what HTM considers the foun- where weights that describe the relationships between units in dations of biological intelligence [15]. It asserts that the sparse the network are learned, however unlike deep learning models 978-1-5386-4185-9/18/$31.00 c 2018 IEEE 4 | P a g e Intelligent Systems Conference 2018 6-7 September 2018 | London, UK where the connections are fixed these models employ online- learning, which enables them to learn concepts by making new connections over time as well as pruning obsolete connections when the relationships are no longer valid.

Fig. 5. Topological model of the nervous system (adapted) [18].

V. RELATIONSHIPS BETWEEN DEEP LEARNINGAND BRAIN ARCHITECTURE Fig. 6. Dynamically programmed proposed deep learning architecture for Six key concepts of Deep Learning in Artificial Intelligence brain connectome (adapted) [22]. (AI) and the functioning of the human brain reflect the evolving merge of the standard model of the mind and AI’s Fourth in our model is emotion. Using feature repre- deep learning tools. sentations such as Convolutional Neural Networks (CNN), First in our model is the concept of brain memory, anal- the brain’s multimodal emotions information such as visual ogous to computer storage. Memory or storage is key to stimuli, past experiences, context, motion, facial expressions the function of both the human brain and in deep learning. can be simulated in computer systems by deep learning neural Memory in the brain, like a deep learning network, stores network models [23]. CNN basic structure depicted in Fig. 7. input data, weighs parameters, and acts on computed data - however dynamically and nomadically the patterns of neurons and synapses accomplish these actions - and machine deep learning uses dynamic RAM (DRAM), static RAM (SRAM) internally and externally, as all computers are designed to save new data to storage in order to function [19]. Second in our model is learning. Both deep learning and the brain learn from their respective datasets. Both of them use the stored data to execute their intelligent actions. Neural networks require frequent access to data to learn from data and stores the entire dataset in computer memory, just as the brain stores the “brain dataset” of pattern recognizers in the hippocampus, learning from the frequency of the high-level features from cortical neurons [7].

Third and fundamental is the circuit diagram that represents Fig. 7. Basic structure of an convolutional neural network (adapted) [5]. the electrical foundations of the current electronic comput- ers and its similarity to the biological connectome structure Behavior is the fifth concept in our model. The hypothesis [20] that is the focus of intense research investment by the of the integration of Deep Learning and Neuroscience behavior United States government [21]. The complex brain connectome says that 1) neurons can gradually mature their synapses, 2) structure of two connected brain neurons (each of which can neurons in different brain areas can optimize different sensors be connected to up to 10,000 other neurons) can be built to improve over time, and 3) different brain areas are pre- using deep learning architecture. To predict the connectome structured to solve identical computational problems posed between the brain neurons the architecture is built using by behavior. This hypothesis is supported by simulations of convolutional layers, max-pooling layers and recurrent layers the implementation of multiple layers of neurons from neural from the bottom up and a dynamically programmed layer on circuitry deep learning [8]. top to align the output sequences of salient temporal patterns from the two recurrent layers as depicted in Fig. 6 [22]. The sixth concept in the model has the potential to unify all 978-1-5386-4185-9/18/$31.00 c 2018 IEEE 5 | P a g e Intelligent Systems Conference 2018 6-7 September 2018 | London, UK the other concepts, which is the China Brain Project [24] that is [4] J. E. Laird, A. Newell, and P. S. Rosenbloom, “Soar: An architecture the work in China that competes with the Human Connectome for general intelligence,” Artificial intelligence, vol. 33, no. 1, pp. 1–64, Project of the U.S. and also explores the broad relationship 1987. between deep learning and the human brain. The aim of the [5] R. Kurzweil, How to Create a Mind: The Secret of Human Thought Revealed. New York, NY, USA: Penguin Books, 2013. China Brain Project includes creating an artificial brain, to [6] H. Markram, K. Meier, T. Lippert, S. Grillner, R. Frackowiak, S. De- better understand the working of the brain at different levels, haene, A. Knoll, H. Sompolinsky, K. Verstreken, J. DeFelipe, S. Grant, and promote deep learning projects with the collaborative J. P. Changeux, and A. Sariam, “Introducing the Human Brain Project,” assistance of neuroscientists working with AI researchers. The in Procedia Comput. Sci., vol. 7, 2011, pp. 39–42. goals of this project are cognitive computational models, such [7] A. Fontana, “A deep learning-inspired model of the hippocampus as those that will support the emerging standard brain model, as storage device of the brain extended dataset,” arXiv preprint and brain-inspired chips. arXiv:1706.05932, 2017. [8] A. H. Marblestone, G. Wayne, and K. P. Kording, “Toward an inte- gration of deep learning and neuroscience,” Frontiers in computational VI.FUTURE WORK neuroscience, vol. 10, 2016. Our current focus is on exploring the storage function of [9] O. E. Dictionary, “Oxford english dictionary online,” 2007. the human brain, by developing and testing models based [10] I. Opris and M. F. Casanova, “Prefrontal cortical minicolumn: From executive control to disrupted cognitive processing,” pp. 1863–1875, some of the theories for creating artificial minds that are 2014. discussed in this paper. Specifically, we will work on a storage [11] I. Opris, L. Santos, G. A. Gerhardt, D. Song, T. W. Berger, R. E. model based on Rosenblatt’s experimental model for storage Hampson, and S. A. Deadwyler, “Prefrontal cortical microcircuits bind in neural networks [25]. It is our hope that this model will perception to executive control,” Scientific reports, vol. 3, 2013. enable us to replicate some human sensory perceptions such [12] R. Jolivet, J. S. Coggan, I. Allaman, and P. J. Magistretti, “Multi- as hearing, vision, and language as explained with the HTM timescale Modeling of Activity-Dependent Metabolic Coupling in the model discussed in IV. Neuron-Glia-Vasculature Ensemble,” PLoS Comput. Biol., vol. 11, no. 2, 2015. [13] H. Markram, “The blue brain project,” Nature Reviews Neuroscience, VII.CONCLUSION vol. 7, no. 2, pp. 153–160, 2006. We are engaged in a rapidly accelerating area of research [14] S. Fine, Y. Singer, and N. Tishby, “The hierarchical hidden markov that is deeply fundamental to our understanding of the design model: Analysis and applications,” Machine learning, vol. 32, no. 1, pp. 41–62, 1998. of the brain and how it implements object and abstract concept [15] A. Hawkins, J. and Ahmad, S. and Purdy, S. and Lavin, “Biological recognition, structural self-organization, and simplification of and Machine Intelligence (BAMI),” 2016, initial online release complex data. If we can incorporate these understandings into 0.4. [Online]. Available: http://numenta.com/biological-and-machine- a standard model of the mind, it promises to accelerate our intelligence/ capacity for creating intelligent agents that can improve or [16] F. Park, “The Hierarchical Hidden Markov Model : Analysis and even expand the design of biology. This area is engaging Applications,” Machine Learning, vol. 62, pp. 41–62, 1998. significant state-sponsored investment, both in the China Brain [17] D. George and J. Hawkins, “A Hierarchical Bayesian model of invariant Project in China and also in the Human Connectome Project in pattern recognition in the visual cortex,” Proc. 2005 IEEE Int. Jt. Conf. Neural Networks, IJCNN’05., pp. 1812–1817, 2005. the US, which will continue to fuel research. The promise of [18] F. Rosenblatt, “Principles of neurodynamics: perceptrons and the theory further exploration is bringing in greater collaboration across of brain mechanisms.” 1962. the academic disciplines that illuminate our understanding of [19] S. Nahal, The Relationship Between Deep Learning and Brain Function: the brain: artificial intelligence, linguistics, cognitive science, Proceedings of Student-Faculty Research Day. Pleasantville, NY, USA: and neuroscience. Establishing a standard model of the mind CSIS, Pace University, 2017, vol. May 5. is an important part of this collaborative effort and it is [20] “National institute of health human con- our hope that the research community will consider theories nectome project,” 2017. [Online]. Available: that represent the computer science perspective for building https://www.neuroscienceblueprint.nih.gov/connectome/ artificial minds, theories such as those presented in works of [21] D. C. Van Essen, S. M. Smith, D. M. Barch, T. E. Behrens, E. Yacoub, K. Ugurbil, W.-M. H. Consortium et al., “The wu-minn human con- Rosenblatt, Kurzweil, and Hawkins. nectome project: an overview,” Neuroimage, vol. 80, pp. 62–79, 2013. [22] V. Veeriah, R. Durvasula, and G.-J. Qi, “Deep learning architecture with REFERENCES dynamically programmed layers for brain connectome prediction,” in Proceedings of the 21th ACM SIGKDD International Conference on [1] J. E, C. L. Lebiere, and P. S. Rosenbloom, “A Standard Model of the Knowledge Discovery and Data Mining. ACM, 2015, pp. 1205–1214. Mind: Toward a Common Computational Framework across Artificial Intelligence, Cognitive Science, Neuroscience, and Robotics,” to appear [23] M. Ghayoumi and A. K. Bansal, “Multimodal architecture for emotion in AI Magazine, 2017. in robots using deep learning,” in Future Technologies Conference (FTC). IEEE, 2016, pp. 901–907. [2] J. R. Anderson, D. Bothell, M. D. Byrne, S. Douglass, C. Lebiere, and Y. Qin, “An integrated theory of the mind.” Psychological review, vol. [24] M.-m. Poo, J.-l. Du, N. Y. Ip, Z.-Q. Xiong, B. Xu, and T. Tan, “China 111, no. 4, p. 1036, 2004. brain project: basic neuroscience, brain diseases, and brain-inspired computing,” Neuron, vol. 92, no. 3, pp. 591–596, 2016. [3] P. S. Rosenbloom, A. Demski, and V. Ustun, “The sigma cognitive ar- chitecture and system: Towards functionally elegant grand unification,” [25] F. Rosenblatt, “A model for experiential storage in neural networks,” Journal of Artificial General Intelligence, vol. 7, no. 1, pp. 1–103, 2016. Computer and information sciences. Washington, DC: Spartan, 1964.

978-1-5386-4185-9/18/$31.00 c 2018 IEEE 6 | P a g e