Ebook: Biologically Inspired Cognitive Architectures 2010

This volume documents the proceedings of the First International Conference on Biologically Inspired Cognitive Architectures (BICA 2010), which is also the First Annual Meeting of the BICA Society. This conference was preceded by 2008 and 2009 AAAI Fall Symposia on BICA that were similar in content (indeed, the special issue of the International Journal of Machine Consciousness

A.V. Samsonovich (guest editor): International Journal of Machine Consciousness, special issue on Biologically Inspired Cognitive Architectures, Vol. 2, No. 2, 2010.

Like the 2008 and 2009 BICA Symposia, the present BICA 2010 conference contains a wide variety of ideas and approaches, all centered around the theme of understanding how to create general-purpose humanlike artificial intelligence using inspirations from studies of the brain and the mind. BICA is no modest pursuit: the long-term goals are no less than understanding how human and animal brains work, and creating artificial intelligences with comparable or greater functionality. But, in addition to these long-term goals, BICA research is also yielding interesting and practical research results right now.

A cognitive architecture, broadly speaking, is a computational framework for the design of intelligent and even conscious agents. Cognitive architectures may draw their inspiration from many sources, including pure mathematics or physics or abstract theories of cognition. A biologically inspired cognitive architecture (BICA), in particular, is one that incorporates formal mechanisms from computational models of human and animal cognition, drawn from cognitive science or neuroscience. The appeal of the BICA approach should be obvious: currently human and animal brains provide the only physical examples of the level of robustness, flexibility, scalability and consciousness that we want to achieve in artificial intelligence. So it makes sense to learn from them regarding cognitive architectures: both for research aimed at closely replicating human or animal intelligence, and also for research aimed at creating and using human-level artificial intelligence more broadly.

Research on the BICA approach to intelligent agents has focused on several different goals. Some BICA projects have a primary goal of accurately simulating human behavior, either for purely scientific reasons – to understand how the human mind works – or for applications in domains such as entertainment, education, military training, and the like. Others are concerned with even deeper correspondence between models and the human brain, going down to neuronal and sub-neuronal level. The goal in this approach is to understand how the brain works. Yet another approach is concerned with designing artificial systems that are successful, efficient, and robust at performing cognitive tasks that today only humans can perform, tasks that are important for practical applications in the human society and require interaction with humans. Finally, there are BICA projects aimed broadly at creating generally intelligent software systems, without focus on any one application area, but also without a goal of closely simulating human behavior. All four of these goals are represented in the various papers contained in this volume.

The term BICA was coined in 2005 by Defense Advanced Research Projects Agency (DARPA), when it was used as the name of a DARPA program administered by the Information Processing Technology Office (IPTO). The DARPA BICA program was terminated in 2006 (more details are available at the DARPA BICA web page at http://www.darpa.mil/ipto/programs/bica/bica.asp). Our usage of the term “BICA” is similar to its usage in the DARPA program; however, the specific ideas and theoretical paradigms presented in the papers here include many directions not encompassed by DARPA's vision at that time. Moreover, there is no connection between DARPA and the BICA Society.

One of the more notable aspects of the BICA approach is its cross-disciplinary nature. The human mind and brain are not architected based on the disciplinary boundaries of modern science, and to understand them almost surely requires rich integration of ideas from multiple fields including computer science, biology, psychology and mathematics. The papers in this volume reflect this cross-disciplinarity in many ways.

Another notable aspect of BICA is its integrative nature. A well-thought BICA has a certain holistic integrity to it, but also generally contains multiple subsystems, which may in some cases be incorporated into different BICAs, or used in different ways than the subsystem's creator envisioned. Thus, the reader who is developing their own approach to cognitive architectures may find many insights in the papers contained here useful for inspiring their own work or even importing into their own architecture, directly or in modified form.

Finally we would like to call attention to the relationship between cognition, embodiment and development. In our view, to create a BICA with human-level general intelligence, it may not be necessary to engineer all the relevant subsystems in their mature and complete form. Rather, it may be sufficient to understand the mechanisms of cognitive growth in a relatively simple form, and then let the mature forms arise via an adaptive developmental process. In this approach, one key goal of BICA research becomes understanding what the key cognitive subsystems are, how do they develop, and how they become adaptively integrated in a physical or virtual situated agent able to perform tight interactions within its own body, the other entities and the surrounding environment. With this sort of understanding in hand, it might well be possible to create a BICA with human-level general learning capability, and teach it like a child. Potentially, a population of such learners could ignite a cognitive chain reaction of learning from each other and from common resources, such as human instructors or the Internet.

Currently BICA research is still at an early stage, and the practical capabilities of BICA systems are relatively undeveloped. Furthermore, the relationships between the ideas of various researchers in the field are not always clear; and there is considerable knowledge in relevant disciplines that is not yet fully incorporated into our concrete BICA designs. But BICA research is also rapidly developing, with each year bringing significant new insights, moving us closer to our ambitious goals. In this sense, the newborn BICA society, according to the intentions of the Founding Members, will be a main vehicle for the growth and dissemination of breakthrough research in the field of BICA systems. The papers presented in this volume form part of this ongoing process, as will the papers in the ongoing BICA conferences to follow.

Alexei V. Samsonovich, Kamilla R. Jóhannsdóttir, Antonio Chella and Ben Goertzel

Editors

The vision system is perhaps the most well understood part of the neocortex. The input from the eyes consists of a set of images made up of pixels that are densely packed in the fovea and less so in the periphery. Each pixel is represented by a vector of attributes such as color, brightness, spatial and temporal derivatives. Pixels from each eye are registered in the lateral geniculate nucleus and projected to the cortex where they are processed by a hierarchy of array processors that detect features and patterns and compute their attributes, state, and relationships. These array processors consist of Cortical Computational Units (CCUs) made up of cortical hypercolumns and their underlying thalamic and other subcortical nuclei. Each CCU is capable of performing complex computational functions and communicating with other CCUs at the same and higher and lower levels. The entire visual processing hierarchy generates a rich, colorful, dynamic internal representation that is consciously perceived to be external reality. It is suggested that it may be possible to reverse engineer the human vision system in the near future [1].

Deep-layer machine learning architectures continue to emerge as a promising biologically-inspired framework for achieving scalable perception in artificial agents. State inference is a consequence of robust perception, allowing the agent to interpret the environment with which it interacts and map such interpretation to desirable actions. However, in existing deep learning schemes, the perception process is guided purely by spatial regularities in the observations, with no feedback provided from the target application (e.g. classification, control). In this paper, we propose a simple yet powerful feedback mechanism, based on adjusting the sample presentation distribution, which guides the perception model in allocating resources for patterns observed. As a result, a much more focused state inference can be achieved leading to greater accuracy and overall performance. The proposed paradigm is demonstrated on a small-scale yet complex image recognition task, clearly illustrating the advantage of incorporating feedback in a deep-learning based cognitive architecture.

We present a universal building block for cognitive machines, called NeuroNavigator, inspired by theories of the hippocampus. The module is designed to fit both biological plausibility and constraints of forthcoming neuromorphic hardware. Its functions may range from spatial navigation to episodic memory retrieval. The goal of the present study of NeuroNavigator is to show the scalability of the model. The core of the architecture is based on our previously described model of hippocampal function and includes 3 layers (DG, CA3, CA1) of spiking neurons with noisy STDP synaptic connections among neighboring layers. The model is applied to a spatial navigation paradigm in a hierarchical virtual environment, the metrics of which need to be learned by exploration. The goal in each trial is set arbitrarily as any one of the previously seen objects or features. In order to navigate toward the goal, the agent needs to “imagine” previously performed available moves at the current location and select one of them, using the acquired spatial knowledge. This process controlled by NeuroNavigator is repeated until the goal is reached. Overall, the simulation results show robustness and scalability of the solution based on a biologically-inspired network of spiking neurons and STDP synapses.

Our goal is to lay a foundation for a model of high level cognition in humans that respects biological constraints. To do so, we integrate realistic neural models of attention, memory and control into an active information foraging system comprising many of the component neural systems and strategies in published models of high-level active vision. The direction of attentional focus requires integration of bottom-up and top-down attention. We base our modeling methodology for both individual components and the overall system operation on structured hierarchical Bayesian models. These models are built upon computationally tractable, neurally implementable sampling approximations to inference and control.

Two related questions posed to biology-inspired cognitive robotics are: (1) what form might a motor plan take in the brain, and (2) how might such a plan be converted into coordinated motor behavior? One approach to modeling the motor plan is to assume a stream of discrete instructions. This conceptualization has been attractive in cognitive science because it allows theorists to discuss motor planning (and perception) in terms of notation systems. However, many brain and robotics researchers have come to view the idea of the discrete motor plan to be wanting. SCA-Net provides a concrete alternative based on established brain circuitry. The approach allows for the coding of perception and production representations in a way that is useful to notation-oriented cognitive architectures.

Human heading perception is robust in the presence of moving objects, except when the object crosses the focus of expansion (FoE). If the object approaches the observer, heading perception is shifted towards the object FoE [1]. If the object maintains a fixed distance, heading perception is shifted away from the object FoE [2]. This data has resulted in the theory that differential motion operators are used to determine heading. Primate neurophysiology, however, indicates that additive MT cells which combine data across pools of motion sensitive cells in V1, and not differential cells, are used to assess heading in dorsal MST [3]. The present work utilizes components of the ViSTARS (Visually-guided Steering, Tracking, and Route Selection) Biologically Inspired Cognitive Architecture (BICA) to analyze and reconcile the data.

ViSTARS BICA components relating to heading perception [4] were analyzed and minimal modifications were made to provide distance dependent weighting in the template match. The results indicate that differential motion operators are not required to explain human heading bias, provided that a detailed model of V1-MT-MSTd is used. The ViSTARS BICA indicates that heading bias in the presence of moving objects results from the representation of heading as a template and the use of a population codes to represent spatial properties within the system. Work is ongoing to fully characterize the causes and nature of this bias. This work demonstrates the utility of BICA not only in extending our computational capabilities but also in providing an overall framework with which to understand and reconcile the data on which it is built.

In this work we present a biologically motivated framework for the modelling of the visual scene exploration preference. We aim at capturing the statistical patterns that are elicited by the subjective visual selection and reproduce them via a computational system. The low level visual features are encoded through the projection of the image patches on a learned basis of linear filters reproducing the typical response properties of the primary visual cortex (V1) receptive fields of mammals. The resulting training set is typically high-dimensional and sparse. We exploit the sparse structure by clustering together patterns of channel activation which are similar on the basis of a binary activation map and finally deriving a pooling over the set of the original linear filters in terms of active (on) and non-active (off) channels for each cluster. The system has been tested on a dataset of natural images by comparing the fixation density maps recorded from human subjects observing the pictures and the saliency maps computed by our system obtaining promising results.

In this paper we illustrate the cognitive architecture of a humanoid robot based on the proposed paradigm of Latent Semantic Analysis (LSA). The LSA approach allows the creation and the use of a data driven high-dimensional conceptual space. This paradigm is a step towards the simulation of an emotional behavior of a robot interacting with humans. The Architecture is organized in three main areas: Sub-conceptual, Emotional and Behavioral. The first area processes perceptual data coming from the sensors. The second area is the “conceptual space of emotional states” which constitutes the sub-symbolic representation of emotions. The last area activates a latent semantic behavior related to the humanoid emotional state. The robot generates its overall behavior also taking into account its “personality”. To validate the system, we implemented the system on a Aldebaran NAO humanoid robot.

The evolutionary thread leading to increasingly more sophisticated types of movement is examined. The entire thread is seen in humans and includes: collision avoidance, landmark navigation, moving to attain good expectations or avoid bad expectations, and moving to actualize an imagined world state. These behaviors compete for the same sensory and motor resources; each requiring greater perceptual discrimination, more complex internal models of the world, and a longer time frame than the previous behavior. By finding the brain structures supporting these complex movement tasks and how tasks compete for resources, researchers will facilitate both our understanding of the brain and our ability to create artificial minds that we understand intuitively.

In this position paper we argue that BICA must simultaneously be compatible with the explanation of human cognition and support the human design of artificial cognitive systems. Most cognitive neuroscience models fail to provide a basis for implementation because they neglect necessary levels of functional organisation in jumping directly from physical phenomena to cognitive behaviour. Of those models that do attempt to include the intervening levels, most either fail to implement the required cognitive functionality or do not scale adequately. We argue that these problems of functionality and scaling arise because of identifying computational entities with physical resources such as neurons and synapses. This issue can be avoided by introducing appropriate virtual machines. We propose a tool stack that introduces such virtual machines and supports design of cognitive architectures by simplifying the design task through vertical modularity.

The Experimental Cognitive Robot version 1 (XCR-1) is a simple three-wheel platform for implementation experiments with the “Haikonen cognitive architecture”. This platform is non-digital and non-programmable; it is not based on any microprocessor. It is not based on pre-programmed algorithms or any kind of programs. Instead, it is based on associative neuron groups in a widely cross-connected architecture where information is carried by spatial and temporal patterns of neural signals. No common code for these signal patterns is used or is necessary. The Haikonen cognitive architecture is a multi-modal perceptive system that utilizes so called perception/response feedback loops for each sensory modality. These modalities are associatively cross-connected. Therefore any experiment with the architecture would have to include multiple sensory modalities so that the cross-connection effects could be demonstrated. Accordingly the robot XCR-1 has sensory modalities for vision, sound, touch, shock (“pain”) and “pleasure”. The robot is able to move with two wheel motors and it also has a gripper mechanism, which can grab suitable-sized objects. In addition to the motor responses the robot has audible “self-talk”, a verbal report, which reflects the robot's “mental perception” of the on-going situation. Presently no associative neuron chips are available. Therefore, in order to minimize the required hardware, the modalities are realized in the most minimal way. However, the realizations are sufficient for the demonstration of many essential issues.

Humans are intrinsically motivated to learn. Such motivation is necessary to be a human-like learner, and helpful for any learning system designed to achieve general intelligence. We discuss the limited existing computational work in this area, and link them to known and theorized properties of the dopamine system. The relatively well-understood mechanisms by which dopamine release signals unpredicted reward can also serve to signal new learning. Dopamine release leads to maintenance of current representations, which serves to “lock” attention onto topics or tasks in which useful learning is occurring. We thus propose a novel but natural extension of known aspects of dopamine function to perform self-directed learning of arbitrary self-defined tasks. If this hypothesis is correct, detailed experimental evidence on dopamine function can help guide computational research into human-like learning systems.

This paper investigates issues of robot's personalization and long-term adaptation in human-robot interaction (HRI). It demonstrates the design and first technical implementation of a HRI showcase in the Robot House at University of Hertfordshire, UK. Here the central idea facilitating the long-term HRI is the creation of robotic companion, which provides various types of service to the user and can be personalised based upon individual needs. The personalisation can also be further enhanced through repeated interactions. The key component in the “mind” of the companion, which is highlighted in this paper, is the model of human semantic and episodic memory. The memory not only allows the companion to remember user's preferences for practical daily tasks, it also changes companion's behaviour in a longer time scale based on robot's perception of actual user input. Finally, implications of such a memory model in HRI are discussed.

GMU BICA is a biologically inspired cognitive architecture developed at George Mason University. Its main distinguishing feature is a system of data structures called “mental states” that enables various forms of metacognition. The present study develops an understanding of the role of metacognition during working scenario generation (a general element of the cognitive cycle in GMU BICA). This is done by a computational experiment with a rapid prototype of the architecture that generates metacognitive mental states. A spatial learning / spatial navigation paradigm is used here, with a virtual robot implemented in CASTLE. Results show a significant increase in both behavioral and cognitive efficiency of the agent when metacognition is turned on. Generalizations are discussed.

Trust is a vital component of human society. Without trust, we would all have to continuously focus on survival, i.e., our physical security and the daily finding of food and water. With trust, we developed the division of labor that allowed us to sleep safely and develop a complex society. However, the concept of trust is very defused and not described operationally, i.e., in terms of the mechanisms by which we build, maintain, and decide to trust. There is now evidence that trust has a biological foundation [1]. That brain imaging data shows that some of the regions of our brain associated with trust are part of our high-level reasoning but others are deep within the brain. Like emotions, deep brain processes are typically fast, unconscious, and unexplainable. My hypothesis is that trust is based on cognitive mechanisms, likely the brain's slow and fast processes [2]. This talk focuses on a computational cognitive modeling approach to developing and understanding the mechanisms of trust in humans. The approach to investigating the mechanisms of trust is the use of computational cognitive modeling, specifically ACT-R [3]. Cognitive modeling is the development of declarative and procedural knowledge that “runs” on a cognitive architecture and produces outputs/behaviors comparable to those of humans. The result of this work will be an ACT-R model replicating human behavior on trust and explaining possible mechanisms of trust. However, the ACT-R system is intended to model rational behavior. Therefore, I will be also discussing mechanisms not normally considered to be “rational” [2]. Knowing the mechanisms associated with trust will allow a principled approach to developing intelligent systems we can and will trust.

Many biological agents are characterized by consciousness both as a capacity to feel as a style of cognitive access to integrated information. It is plausible to assume that some essential feature of cognition has not yet been addressed although it is implemented in many biological agents. Among the aspects of biological cognition underrepresented in the literature on cognitive agents it is worth to mention the double layered fringe-nucleus model proposed by William James. This model outlines the relation between what is already in the focus of attention and what is not. In the recent past it has already been suggested to use James's fringe as a viable model for cognition particularly by Bruce Mangan. Here, the main idea is that the fringe-nucleus model is not limited to the cognitive modeling but rather it shapes the relation between the conscious mind and the environment. The fringe zone could model the transition area between what is integrated inside the cognitive domain and the environment. In this paper, we explore the link between the fringe-nucleus model and an externalist, situated and embodied framework whereby the vehicles of cognition are spread so to comprehend the external world.

ATMAN in Sanskrit means “The Self”. ATMAN is an architecture of the mind developed by observing the different anomalies of the brain and hypothesizing a system that would display such behavior. This paper introduces a system demo for the use of “ATMAN” model to play a game of tic-tac-toe and some of its side effect discoveries.

The segmentation of a stream of perceptual inputs a robot receives into discrete and meaningful events poses challenge in bridging the gap between internal cognitive representations, and the external world. Event Segmentation Theory, recently proposed in the context of cognitive systems research, sustains that humans segment time into events based on matching perceptual input with predictions. In this paper we propose a framework for online event segmentation, targeting robots endowed with active perception. Moreover, sensory processing systems have an intrinsic latency, resulting from many factors such as sampling rate, and computational processing, and which is seldom accounted for. This framework is founded on the theory of dynamical systems synchronization, where the system considered includes both the robot and the world coupled (strong anticipation). An adaption rule is used to perform simultaneous system identification and synchronization, and anticipating synchronization is employed to predict the short-term system evolution. This prediction allows for an appropriate control of the robot actuation. Event boundaries are detected once synchronization is lost (sudden increase of the prediction error). An experimental proof of concept of the proposed framework is presented, together with some preliminary results corroborating the approach.

This paper briefly describes four kinds of learning carried out by intelligent agents in a computational environment facilitating joint activities of people and software agents. The types of learning and the applications we draw examples from are: learning by being told and learning by experience, as illustrated through a virtual patient application; learning by reasoning, as illustrated through a clinician's advisor application; and learning by reading, as illustrated by an ontology enhancement application. The agents carrying out these types of learning are modeled using cognitive modeling strategies that show marked parallels with how humans seem to learn.

The ASMO Cognitive Architecture has been developed to support key capabilities: attention, awareness and self-modification. In this paper we describe the underlying attention model in ASMO. The ASMO Cognitive Architecture is inspired by a biological attention theory, and offers a mechanism for directing and creating behaviours, beliefs, anticipation, discovery, expectations and changes in a complex system. Thus, our attention based architecture provides an elegant solution to the problem of behaviour development and behaviour selection particularly when the behaviours are mutually incompatible.

Emergent behaviours are believed to be a property of complex cognitive architectures. However it is difficult to observe them in carefully engineered systems. We believe that, in order to obtain interesting and surprising behaviours, the designer has to remove some predictability requirements.

To explore this we studied how the complexity, defined according to the Kolmogorov theory, is linked to truly emergent behaviours. This study has been conducted on a real robot driven by the non-linear dynamics of a recurrent neural network created using evolutionary approaches. In several experiments we found that, while a simple network is equivalent to a deterministic automata, networks of increasing complexity exhibit novel behaviours. This is due to a spontaneous background activity of the neurons which can be sustained only by a complex network.

These results support the idea that in a non predictable architecture novel and surprising cognitive capabilities emerge in a natural way.

Initially developed to support reflective reasoning in computer vision, over the past ten years GRAVA has been applied to a wide range of problems including vision, language, networks, cyber security, and neural modeling. In this paper we provide an overview of the key concepts in GRAVA, how they have evolved over the ten year history, and how they have been applied.

Graphical cognitive architectures implement their functionality through localized message passing among computationally limited nodes. First-order variables – particularly universally quantified ones – while critical for some potential architectural mechanisms, can be quite difficult to implement in such architectures. A new implementation strategy based on message decomposition in graphical models is presented that yields tractability while preserving key symmetries in the graphs concerning how quantified variables are represented and how symbols, probabilities and signals are processed.