Ebook: Biologically Inspired Cognitive Architectures 2011

This volume documents the proceedings of the Annual International Conference on Biologically Inspired Cognitive Architectures BICA 2011, which is the Second Annual Meeting of the BICA Society, and the fourth annual BICA meeting. Like in the previous years, the main body of the present volume contains a wide variety of ideas and approaches documented in research papers, position papers and abstracts presented at the conference that are all sorted alphabetically by the last name of the first author, with the exception of the first introductory chapter (intended as a guide to the Reader). The main body is followed by two supplementary parts. All papers and abstracts included in this volume were carefully peer-refereed by the Program Committee members and external reviewers (the acceptance rate was 83%).

Throughout this book, the acronym “BICA” stands for Biologically Inspired Cognitive Architectures: computational models that incorporate formal mechanisms of human and animal cognition, drawn from cognitive science and neuroscience into artificial intelligence (for a definition, please see the first and the last chapters). The series of BICA conferences was initiated in 2008 under the umbrella of AAAI Fall Symposia Series, as a follow-up on the abruptly terminated DARPA BICA program. Over the last 4 years, BICA conference developed into a mid-size international conference organized and sponsored by the BICA Society. The conference focuses on emergent hot topics in computer and brain sciences unified by the challenge of replicating the human mind in a computer.

Given the size of the volume, even a brief overview of its content cannot be given in a short preface. Therefore, for an introduction, we refer the reader to the first chapter (by K.R. Jóhannsdóttir and A.V. Samsonovich) that provides a brief overview of this volume, placing it in the broader context of recent emergent developments in the field of BICA. This introductory chapter also serves as a read-me-first document and a guide to the book. Papers reflected in it are grouped by topics and are connected to each other. Anybody who wants to start learning about the field of BICA by reading this book is encouraged to start with the introductory chapter.

We would like to thank all members of the Organizing and Program Committees for their help with the organization of this conference, selection and review of the materials. In particular, we are very grateful to Drs. Antonio Chella, Brandon Rohrer, and Christian Lebiere for their active service on the Core Organizing Committee. We also wish to thank all those who helped us with local organization and arrangements, including Mr. Michael Galvin, Ms. Grace Radfar, Ms. Nicole Charles, and many others. The last, but not the least, is our great thanks to Chris Poulin for his generous financial support of the discussion panel on the Roadmap to Human-Level AI, made possible via a donation to the BICA Society from Patterns and Predictions.

Alexei V. Samsonovich and Kamilla R. Jóhannsdóttir, Editors

This chapter provides a brief overview of the volume of BICA 2011 Proceedings, viewed in the broader context of recent emergent developments in the field of BICA. The present conference is the fourth event in the BICA conference series, and the Second Annual Meeting of the BICA Society. In addition to being a guide to the book, this chapter provides a general overview of the field of BICA, and is recommended to anybody who wants to learn about the field by reading this book.

Deep generative models and their associated top-down architecture are gaining popularity in neuroscience and computer vision. In this paper we link our previous work with regulatory feedback networks to generative models. We show that generative model's and regulatory feedback model's equations can share the same fixed points. Thus, phenomena observed using regulatory feedback can also apply to generative models. This suggests that generative models can also be developed to identify mixtures of patterns, address problems associated with binding, and display the ability to estimate numerosity.

Due to recent research, the neurobiological elements behind mental disorders such as depression become more and more clear. This paper presents an integrated computational model based on neurobiological insights and psychological theories. The model is used to analyse the effect of existing psychological treatments. The simulation experiments give an insight in the interaction between different cognitive components in mental disorders and illustrates why different treatments can have different effects for people with different genetic dispositions.

Building robots that closely resemble humans allows us to study phenomena that cannot be studied using mechanical-looking robots in our daily human-to-human natural interactions. This is supported by the fact that human-like devices can more easily elicit the same kind of responses that people use in their natural interactions. However, several studies support the close and complex relationship existing between outer appearance and the behavior by the robot. Yet, human-like appearance, as Masahiro Mori observed, is not enough to give a positive impression. The robot has to behave closely to humans, and is to have a sense of perception that enables it to communicate with humans. Our past experience with android “Geminoid HI-1” demonstrated that the sensors equipping the robot are not enough to perform human-like communication, mainly because of a limited sensing range. To overcome this problem, we endowed the environment around the robot with perceptive capabilities by embedding sensors such as cameras into it. This paper reports a preliminary study about the improvement of the controlling system by integrating cameras in the surrounding environment, so that the android ca be given human-like perception. The integration of the development of androids and the investigations of human behaviors constitute a new research area merging engineering and cognitive sciences.

Relational mapping, a cognitive sub-process of relational reasoning, plays a critical role in identifying similarities between abstract constructs. This paper discusses an initial endeavor in developing a biologically-inspired spiking neuron model that performs relational mapping in a similar functional manner to existing cognitive models founded in neuroscience. Using spiking neurons provides a capability to portray neural dynamics that naturally lead to notions of critical relational mapping sub-functions such as binding by synchrony. The model, although still in progress, is a step in the direction of progressing cognitive concepts down to an individual spiking neuron level.

This paper presents the development of a prototype cognitive system for tutoring pediatric telephone triage to resident doctors. The cognitive system is built through a spiral development methodology by using the Disciple learning agent shell. This illustrates solutions to challenges encountered in developing cognitive systems in new domains with a low budget, including prioritizing the development, reuse of general purpose modules, and experimentation-driven development.

Prefrontal cortex (PFC) is implicated in a number of functions including working memory and categorization. Here the Prefrontal cortex Basal Ganglia Working Memory (PBWM) model (O'Reilly and Frank, 2006) is applied to learning categories with invariances. In particular, motivated by a problem in scene recognition, objects in different locations are sequentially presented to the network for categorization. The model learns to recognize these classes without explicit programming, thus modeling human categorization along with characteristics such as generalization to novel sequences and frequency dependent effects. Future extensions to the current work including applications to other domains and modeling functionally distinct segregations of PFC and neuromodulatory systems are also described.

Epigenetic and enactive robotics have been proposed as test-beds for psycho-biological models of the mind. These approaches shortened the distance between the artificial and the natural mind by stressing the importance of the unity between the brain, the body and the environment. At the same time, nowadays robotic researchers openly acknowledge the importance of experience which has been not sufficiently considered in the recent past. The process went so far that the field of machine consciousness is now part of the scientific landscape.

The externalist approach identifies experience with relations, processes or acts between an agent and its environment. Externalism is the view according to which the brain and its neural activity is necessary but not sufficient to produce the conscious mind. The externalist approach locates the subject and experience processes in a context wider than brain-oriented approaches. Because of this fact, the externalist standpoints allows to start from methodological and ontological premises suitable for the study of experience and of other subjective contents inside an experimental framework.

The talk will review the externalist-oriented approaches and it will present the main ideas at the basis of the emerging field of externalist robotics. The aims of this research is to experiment, in a circumscribed number of cases, whether it is possible to apply such an architecture in the fields of robotics, of psychology and of philosophy of mind.

In this paper we illustrate a new version of the cognitive architecture of an emotional humanoid robot based on the proposed paradigm of Latent Semantic Behaviour (LSB). This paradigm is a step towards the simulation of an emotional behavior of a robot interacting with humans. The New Architecture uses a different procedure of induction of the emotional conceptual space and an Android mobile phone as user-friendly for the emotional interaction with robot. The robot generates its overall behavior also taking into account its “personality” encoded in the emotional conceptual space. To validate the system, we implemented the distribute system on a Aldebaran NAO humanoid robot and on a Android Phone HTC and we tested this new emotional interaction between human and robot through the use of a phone.

A step is taken towards fusing symbolic and decision-theoretic problem solving in a cognitive architecture by implementing the latter in an architecture within which the former has already been demonstrated. The graphical models upon which the architecture is based enable a uniform implementation of both varieties of problem solving. They also enable a uniform combination with forms of decision-relevant perception, highlighting a potential path towards a tight coupling between central cognition and peripheral perception.

This paper explores parallels between Reservoir Computing models and basal ganglia nuclei including the subthalamic nucleus and globus pallidus. The anatomy of these structures can inform the design of reservoirs, possibly improving performance and smoothing the reservoir learning error landscape.

A system which performs a complex combination of behaviours has two superficially independent architectures. One is the functional architecture, which separates the behavioural features of the system into feature modules made up of groups of similar behaviours, and defines the interactions between features. The other is the system architecture (alternatively called the physical or information process architecture) which separates the physical information handling resources of the system into modules that perform different types of information processes, each module optimized to perform a different type of process. Any one feature module will employ information processes performed by many or all resource modules. Many different functional architectures are possible, but the need to limit the resources supporting large numbers of different behaviours tends to constrain the form of the system architecture. In the limiting case as the ratio of the number of behaviours learned to the available resources becomes very large, the system architecture is constrained into a very specific form. In the case of a complex learning system this form is called the recommendation architecture. Because there are natural selection advantages for species that require fewer neural resources to learn a given set of behaviours, there is a tendency for the recommendation architecture form to appear in biological brains including human, mammal and avian brains. A system designed to perform a complex combination of behaviours will be much more effective if designed within this form.

In previous work, we have argued that a sophisticated cognitive system with a complex body must possess configurable models of itself (or at least its body) and the world, along with the necessary infrastructure to use the modelled interactions between these two components to select relatively advantageous actions. These models may be used to generate representations of the future (imagination) and the past (episodic memory). In this paper we will explore some problems surrounding the representation of the present arising from the use of such models in the artificial cognitive system under development within the ECCEROBOT project. There are two aspects to consider: the representation of the state of the robot's body within the self model, and the representation of the state of the external world within the world model. In both natural and robotic systems, the processing of the sensory data carrying state information takes a considerable time, and so any estimates of the present states of both the agent and the world would have to be obtained by using predictive models. However, it appears that there is no need for any such representations to be generated in the course of selecting a course of action using self and world models, since representations are only of the future or the past. This may call into question the utility and timing of the apparent perception of the present in humans.

This article takes a brief look at the history of Artificial Intelligence (AI), from good old-fashioned AI (GOFAI) to situated and embodied AI (SEAI) and its relationship to cognitive incrementalism, wherein sensorimotor mechanisms form the basis for high-level cognition. Artificial neural networks (ANNs) designed and tuned by evolutionary algorithms (EAs) are discussed in terms of their potential contributions to SEAI. Though state-of-the-art evolutionary ANN (EANN) research has not fulfilled this promise, our script-based EANN system (SEVANN) is briefly introduced as a software tool for quickly testing the SEAI utility of neurocomputational models of various spatial and temporal granularities.

We hypothesize that the brain “secrets” constitute a moderate number (n < 100) of system principles at all levels of brain construction – from biomolecular structures up to the higher nervous activity (the latter term for mind functions was used by I.P. Pavlov (1849-1936)). We believe that now an unbiased search (by a specialized team of 21 intensely cooperating explorers) of these principles (many of them are already known) and their live implementations might yield a complete human brain wiring diagram with its full functional description in less than five years. Of course, the teamers should be highly motivated in the subject of the search and well-equipped with modern high-tech research tools and knowledge with the background in physics and informatics. The successful solution of the stated above problems will further enable artificial brain creation. With these ideas, we are starting the Russian project in brain reverse engineering, which is planned to run on full swing at the end of 2012.

Our hopes for success are based on great achievements of several past decades in all branches of neuroscience, and the experience of previous works, which will be summarized in the talk. In particular, we will deal with data formats in neural channels (frequency and ensemble codes, stochastic resonance, synchronous/asynchronous modes of multineuronal activity, stochastic (Marr-Hopfield) and continuous (bump) attractor networks and different types of spike-time dependent plasticity. We are also planning to integrate into our operation many parts of the much underappreciated legacy of David Marr (1946-1980) in understanding cortical neural networks as well as all available relevant theoretical and experimental results of the recent years.

The human mind can do many amazing things. Of particular interest are a set of cognitive abilities such as simple inference and recognition that are computationally very demanding, but that we humans perform without any perceptible delay or any sense of mental effort – this despite the fact that our brains use slow, millisecond-speed components. In this position paper, I present a brief inventory of human mental operations that exhibit this kind of surprising efficiency. I suggest that we humans accomplish these feats by (a) avoiding the computationally intractable forms of these problems and (b) by applying massively parallel processing, though perhaps of a very simple kind. This is not a new suggestion – some of these ideas were first developed in my mid-1970s work on NETL [1]. However, I believe that a renewed focus on the parallel vs. serial components of mental processing can help us both in understanding human intelligence and in achieving human-like performance in our AI systems.

The ability to generate, narrate, and understand stories allows humans to accomplish tasks they would otherwise find difficult or impossible. We draw on observations of human narrative to identify a number of capabilities underlying the narrative faculty that we assert must be integrated into any cognitive architecture intended to achieve human-level performance. In particular, we identify sequencing, gap-filling, and plot pattern extraction as key capabilities, and detail two systems under development in our research group that unify a number of cognitive functions to achieve those abilities, functions including multi-representationalism, dynamically-produced commonsense knowledge, and analogical reasoning. Interestingly, in humans, the narrative faculty is intimately tied to cultural transmission and understanding. We speculate that culture is, in fact, a necessary cognitive resource (akin to commonsense, or the lexicon), and that any cognitive architecture intended to achieve human-level performance will need to have a ‘culture’ of its own.

The detection and recognition of a human face should meet the need for social interaction that drives a humanoid robot, and it should be consistent with its cognitive model and the perceived scene. The paper deals with the description of the potential of having a system of emotional contagion, and proposes a simple implementation of it. An emotional index allows to build a mechanism which tends to align the emotional states of the robot and the human when a specific object is detected in the scene. Pursuing the idea of social interaction based on affect recognition, a first practical application capable of managing the emotional flow is described, involving both conceptual spaces and an emotional/motivational model of cognitive architecture

This paper outlines a proposal for a two-level cognitive architecture reproducing the process of abstract thinking in human beings. The key idea is the use of a level devoted to the extraction of compact representation for basic concepts, with additional syntactic inference carried on at a meta-level, in order to provide generalization. Higher-level concepts are inferred according to a principle of simplicity, consistent with Kolmogorov complexity, and merged back into the lower level in order to widen the underlying knowledge base.

A current project is briefly described, involving the application of the OpenCog integrative AGI system to control an animated agent in a videogame world inspired by the commercial game Minecraft. Among other purposes the project will provide a means of testing the hypothesis that cognitive synergy among learning mechanisms associated with different types of memory is valuable for embodied, human-like general intelligence.

This presentation will detail the inception, development and evaluation of ASKNet. A system which uses natural language processing tools in order to create psycholinguistically inspired, spreading activation based semantic networks from natural language texts.

The only known path to general intelligence is that taken by humans. Adapting elements of this path to achieving artificial general intelligence (AGI) has become a common area of interest. We address the role of human teachers in this process, using the concept of the zone of proximal development (ZPD). We explore the range of possible human-teacher interactions, including those modeled closely on humans, those involving accessing and changing the AGI learners internal representations, and tighter integrations amounting to human-AI hybrid learning system (HAIHLS). In such a system, a human teacher scaffolds an untrained subsystem by producing the outputs desired from a fully trained version. Those outputs both train that subsystem and provide more useful information to the remainder of the cognitive system. This aid enables all subsystems to learn within the context of the richer behavior and cognition possible with the aid of the human subsystem.