BICA-2009 Tentative Abstracts

July 23, 2009

Key speakers include, alphabetically: Igor Aleksander (Imperial College), Bernard Baars (NSI), Theodore Berger (USC), Kenneth De Jong (GMU), Stan Franklin (U Memphis), Stephen Grossberg (Boston U), Benjamin Kuipers (U Michigan), Chris Lebiere (CMU), Konstantin Likharev (Stony Brook U), Carol O'Donnell (DOE IES NCER), Jim Reggia (UMD), Frank Ritter (Penn State U), Stuart Shapiro (U Buffalo), Hans-Georg Stork (European Commission), Kristinn Thorisson (CADIA / Reykjavik U)

IMPORTANT NOTE TO THE AUTHORS: A full paper or an (extended) abstract submitted to the symposium should not overlap on more than 50 percent of its content with your preliminary tentative abstract posted here. Posting of the tentative abstract does not guarantee acceptance of your paper and presentation. A posted tentative abstract cannot be a substitute for your submission. Please follow the submission instructions and submit your paper in AAAI format by the deadline via email to samsonovich@cox.net (please see the main page: http://binf.gmu.edu/~asamsono/bica/).

Intended Presentations:

1. Tsvi Achler [Title TBA].

2. Igor Aleksander and David Gamez (keynote). Mental integration experiments with discrete neural networks.

3. Itamar Arel, Derek Rose, and Robert Coop. A biologically-inspired deep learning architecture with application to high-dimensional pattern recognition.

4. Raul Arrabales, Agapito Ledezma, and Araceli Sanchis. Assessing and characterizing the cognitive power of machine consciousness implementations.

5. Bernard J. Baars (invited Memphis talk, title TBA).

6. Gary Berg-Cross. Constructionist epigenetically inspired ideas about cognitive development.

7. Theodore W. Berger (keynote). [title TBA].

8. Vincent Billock and Brian H. Tsou. [Title TBA].

9. Ronald C. Blue. A behavioral developmental psychological comparison of the robots – RICCI and LITTLE BLUE.

10. Piotr Bołtuć et al. The engineering thesis in machine consciousness.

11. Peter G. Burton. A specific model of human cognitive coherence and its control structure: Cognitive System Theory.

12. Antonio Chella and Haris Dindo. How to resolve ambiguities in a grounded human-robot interaction?

13. Vasile Coman. Back to the basics – redefining information, knowledge, intelligence, and artificial intelliegence using only the adaptive systems theory.

14. André Cyr, Mounir Boukadoum and Pierre Poirier. Biologically inspired hebbian learning function for a simulated robot’s artificial spiking neural network controller.

15. Karl Fua, Ian Horswill, Andrew Ortony, and William Revelle. Reinforcement sensitivity theory and cognitive architectures.

16. David Gamez and Igor Aleksander. Taking a mental stance towards artificial systems.

17. Robert S. Gardner, Matteo Mainetti, and Giorgio A. Ascoli. Measuring rates of human memory retrieval.

18. Ben Goertzel. A Deeply-connected hybrid neural-symbolic cognitive architecture.

19. Jaime Gomez and Ricardo Sanz. To cognize is to categorize revisited: Category Theory is where mathematics meets biology.

20. Stephen Grossberg (invited speaker). Autonomous adaptive brain systems and neuromorphic agents.

21. Owen Holland and Hugo Gravato Marques. From functional embodied imagination to episodic memory.

22. Christian Huyck and Emma Byrne. Neural cognitive architectures.

23. Magnus Johnsson, Christian Balkenius, and Germund Hesslow. Neural network architecture for crossmodal activation and perceptual sequences.

24. Benjamin Kuipers (invited speaker). Toward Bootstrap Learning of the Foundations of Commonsense Knowledge.

25. Konstantin K. Likharev (keynote). Biologically inspired computing in CMOL CrossNets.

26. Steve Morphet. An engineering view of cognitive architectures.

27. Carol L. O'Donnell, Elizabeth R. Albro, and Jonathan Levy (invited joint session talk, title TBA).

28. Mikhail I. Rabinovich and Mehmet K. Muezzinoglu. Competitive dynamics of emotion-cognition interaction.

29. James Reggia, Jared Sylvester, Scott Weems, and Mike Bunting (keynote). A simple oscillatory short-term memory model.

30. Frank E. Ritter, Sue Kase, Laura Cousino Klein, Jenette Bennett, and Michael Schoelles (keynote). Using a cognitive architecture to model the effects of stress and caffeine.

31. Brandon Rohrer. Concepts from data I: State-space partitioning with K-D trees.

32. Pablo E. Román and Juan D. Velásquez. Analysis of the web user behavior with a psychologically based diffusion model.

33. Alexei V. Samsonovich. The new schema formalism of Cognitive Constructor: A step beyond the object-oriented paradigm.

34. Ricardo Sanz, Jaime Gomez, Carlos Hernandez and Adolfo Hernando. From brain models to cognitive architectures.

35. Stuart C. Shapiro (keynote). The MGLAIR cognitive architecture.

36. Josefina Sierra. Experiments on the acquisition of linguistic competence in communicating propositional logic sentences.

37. Javier Snaider, Ryan McCall, and Stan Franklin. Time production and representation in a conceptual and computational cognitive model.

38. Andrea Stocco, Christian Lebiere, and John R. Anderson. Dopamine, learning, and production rules: The basal ganglia and the flexible control of information transfer in the brain.

39. Hans-Georg Stork (invited joint session talk). [title TBA].

40. Tyler Streeter and James Oliver. A brain-inspired cognitive architecture for reinforcement learning.

41. Robert Swaine. HELEN: Hierarchical Event-Language Network for Knowledge Representation.

42. Kristinn R. Thorisson (keynote). From constructionist to constructivist AI.

43. Rodrigo Ventura. Emotions: A bridge between nature and society?

44. Pei Wang. Insufficient knowledge and resources - A biological constraint and its functional implications.

45. Mark Waser. A Complete ethical system (and supporting architecture) for intelligent machines.

Tentative Abstracts:

[Title TBA]

Tsvi Achler

Important problems remain where a learning approach is impractical. These problems occur when the environment contains two or more simultaneous patterns (ie. cluttered rainforest, cocktail party conversation). Learning every possible simultaneous pattern combination results in an exponential explosion. How do animals resolve this? How do animals learn the single patterns and are subsequently able to analyze simultaneous patterns? The answer lies in the overwhelming amount of feedback seen throughout sensory processing regions of the brain. Neural Networks containing feedback inhibition can classify simultaneous patterns. I will demonstrate this property by comparing algorithms’ performance (feedback inhibition vs. lateral inhibition, contemporary neural networks and AI algorithms) given inseparable simultaneous patterns.

Mental Integration Experiments with Discrete Neural Networks

(keynote)

Igor Aleksander and David Gamez

In discussions about the physical support of conscious experience, a recent trend has been introduced (largely by Tononi and various colleagues) which measures the ability (given the symbol Φ) of a dynamic network for integrating information. It is argued that, to support conscious experience, the network needs to have highly distinct states that are not separable into the activity of separate processes. The higher the Φ, the better this is said to be. This paper draws attention to the fact that work done in this area has not concerned itself with the phenomenological notion that in conscious organisms the distinct states occur as a result some form of learning or internalization of world events. The paper illustrates, through a series of experiments, what information integration might mean when phenomenal ‘experience’ states shape the state structure of a network. The methodology of weightless generalizing dynamic neural networks is used for this purpose and explained within the paper. The experiments take a ‘mental stance’ through displaying the state of a system so that it can be judged by an observer. Experiments include considering the effect of connectivity in homogenous networks, the effect of localized connections, discontinuities and connections to the external world. We conclude that the maintenance of highly distinct, non-decomposable, phenomenal states needs structural qualifications that go beyond Φ.

A Biologically-Inspired Deep Learning Architecture with Application to High-Dimensional Pattern Recognition

Itamar Arel, Derek Rose, and Robert Coop

Deep learning architectures have received considerable attention during the past few years. They are primarily considered as a scalable tool for processing high-dimensional signals with spatio-temporal dependencies that may span large scales. Recent neuroscience research findings support the notion of hierarchical deep learning as a governing operational paradigm for the neocortex. The goal of artificial deep learning systems is to represent observations in a manner that will later facilitate robust pattern classification - a key attribute of the mammal brain. This stands in contrast with the mainstream approach of pre-processing the data so as to reduce its dimensionality, a practice that often results in sub-optimal performance.

This paper presents a novel, scalable, deep learning architecture which comprises of a basic cortical circuit that homogeneously populates a hierarchical topology. Representation of high-dimensional patterns is achieved in a manner driven by regularities in the observations, such that complex features are inherently formed across the hierarchy. Each node in the proposed architecture maintains a belief state that is shaped by the node’s observation as well as belief information received from upper-level nodes. A Bayesian inference framework is employed, whereby imperfect observations are recoverable through an online clustering algorithm, while state dynamics are learned from experience. Simulation results clearly demonstrate the core capabilities of the proposed framework, particularly in the context of high-dimensional pattern classification.

Assessing and Characterizing the Cognitive Power of Machine Consciousness Implementations

Raul Arrabales, Agapito Ledezma, and Araceli Sanchis

Many aspects can be taken into account in order to assess the power and potential of a cognitive architecture. In this paper we argue that ConsScale, a cognitive scale inspired on the development of consciousness, can be used to characterize and evaluate cognitive architectures from the point of view of the effective integration of their cognitive functionalities. The problem of defining general architectural and behavioral criteria is also discussed in the context of this work.

Constructionist Epigenetically Inspired Ideas about Cognitive Development

Gary Berg-Cross

It also seems useful, if not essential, to understand how cognitive capabilities are acquired developmentally. Thus starts with a reasonable hypothesis that human cognition is genetically seeded but that this genetic component is only one factor in an unfolding network of influences. This paper explores a constructionist, epigenetic understanding of cognitive competencies arising through a prolonged developmental process through which increasingly more complex cognitive structures emerge naturally as embodied cognitive capabilities are coupled to environmental interactions using sensors and effectors (Zlatev & Balkenius 2001). The paper has four parts. First, the rationale for an epigenetic, constructionist view of cognitive development is provided including neurobiological evidence (e.g. largely free of domain-specific structure) for a progressive increase in the representational properties of cortex. The interaction embedded in the real work and neural growth is produces robust type of cognition that minimizes the need for neural pre-specification. Second, the inspiration from these concepts are related the emergence of language in a broad, situated cognitive context including the transforming nature of feedback. An epigenetic equilibration of cognitive structures pulls the development of language along somewhat stable, developmental trajectories. Third, the new interdisciplinary fields of developmental and epigenetic robotics are discussed as ways to investigate particular cognitive models and hypothesis originating from developmental sciences. It is argued that this approach may eventually provide new relevant data about the emergence of cognitive capabilities. Epigenetic robotics provides an alternative path voiding knowledge representations that artificially bottom out in symbolic, programmer-generated descriptions of the world. Epigenetic robotics may be able to explore “neural constructivism” formulations in which the representational features of cortex are built from the dynamic interaction between neural growth mechanisms and environmentally derived neural activity. Finally, some challenges of epigenetic constructionist cognition are entertained. A constructionist epigenetic robotics approach does not assume many of the simplifications of verbal, logic-based descriptions or symbol-processing implementations of cognition. This idea is mitigated by the use of some general principles that might influence cognitive architectures.

[Title TBA]

(keynote)

Theodore W. Berger

[no abstract available at this time] (the keynote will focus on research on neural prosthetics updated with the latest provocative results).

[Title TBA]

Vincent Billock¹ and Brian H. Tsou²

[no abstract available at this time]

A Behavioral Developmental Psychological Comparison of the Robots – RICCI And LITTLE BLUE

Ronald C. Blue

A brief introduction to the historical and theoretical reasons of using the brain model of Correlational Holographic Opponent Processing to electronically form wavelet analog interference memory in two different robots – RICCI and LITTLE BLUE using different procedures. What is important is how a biologically inspired cognitive system adapts or correlates information over time and uses that information at critical moments in a creative way and the observed similarity to human psychological development from an artificial system. Comparisons of the observed development of the two robots will be reported.

http://u2ai.us/ricci.htm

The Engineering Thesis in Machine Consciousness

Piotr Bołtuć et al.

We argue that consciousness can be engineered. This is obviously true of functional consciousness. It is also persuasively put forth in regards to phenomenal consciousness viewed in the framework of cognitive psychology, though there is still much debate about details. Yet, we extend the argument beyond the tradition of behaviorist or reductive views on consciousness that predominate within present-day cognitivism. The argument is that, if Nagel-style non-reductive materialism holds true so that we can understand it, the kind of consciousness it involves can (in principle) be engineered. We draw some preliminary suggestions how this could work.

Keywords: the engineering thesis, machine consciousness, p-consciousness, h-consciousness, the hard problem of consciousness, non-reductive naturalism.

A Specific Model of Human Cognitive Coherence and its Control Structure: Cognitive System Theory

Peter G. Burton

More than a generation ago, David Marr taught us to look in vision for an intermediate form of representation, workable in the real brain, to mediate interpretation of 2d-image data in 3d-depth perception. My Cognitive System Theory seeks to establish a particular kind of intermediate representation, workable in the real brain, that mediates human higher brain function. The CST ‘template’ aggregates sensory information but is managed by system-level principles which operate quite agnostic of template content. These principles liken cognitive capture to refinable investments in repertoire, which investments escalate tasking performance combinatorially. The template model of CST provides an operational basis for relating higher brain function to underlying mechanistic control, in a manner that not only elucidates fundamental cognitive processing in the brain in terms of the concept of punctuated coherence, but also specifies the five modes of conscious control over cognition required in HBF. The template model makes clear why the sub-conscious vs. conscious mode of HBF is necessary, and requires a regulatory and conditioning framework (pure consciousness) operating over the senses to deliver distinct attentional modes for body-management, symbolic thought and navigational tasking. Finally, CST provides a clear insight into the cognitive acquisition of a self-model, and its role in coming to own competences. CST itself is a coherent and encompassing synthesis predicated upon five independent axiomatic analyses: on learning; on experience; on consciousness; on knowledge; and on the self-model: see website repository http://homepage.mac.com/blinkcentral. CST embodies an effective architecture to ensure cognitive resilience through contextually indexed modes of consciousness, which may be useful as a model in the development of more competent artificial cognitive systems.

How to Resolve Ambiguities in a Grounded Human-Robot Interaction?

Antonio Chella and Haris Dindo

The paper describes a trainable system that learns grounded language models from examples with a minimum of user intervention and without feedback. We have focused on the acquisition of grounded meanings of spatial and adjective/noun terms. The system has been used to understand and subsequently to generate appropriate natural language descriptions of real objects, and to engage in verbal interactions with a human partner. Present work also addresses the problem of how to resolve eventual ambiguities arising during verbal interaction through an information theoretic approach.

Back to the Basics – Redefining Information, Knowledge, Intelligence, and Artificial Intelliegence Using Only the Adaptive Systems Theory

Vasile Coman

In a recent interview to the Scientific American, the famous enterprise architect Grady Booch pointed to a "dirty little secret" of today's software. Not only there is no manual how it should be written, but most of the software-intensive systems have an architecture that it's accidental, not intentional. The same lack of directions can be found not only in software, but in almost any information-related field, from economy to ecosystems, from humans to biological systems. This situation is a direct result of lacking answers to some of the most basic questions, such as "what is information" or "what is the knowledge". Without clear definitions that do not require cross-referencing each other, it is very unlikely that much progress will be made, especially in the Biologically-Inspired Cognitive Architecture field.

This paper proposes a return to the system theory, especially the adaptive ones, to redefine all the foundational elements such as information and knowledge. This approach can be successful because both biological organisms and socio-economic organizations can be considered adaptive systems. The new systemic view of information and knowledge not only helps us understand better how our biological brain-body combination works, but also the informational model of our socio-economic organizations. At the end of the paper we intend to present how this new information-centric framework can be applied in practice with great success. Based on these definitions, we introduce new information-centric models for biological organisms and their nervous systems, ecosystems and their self-regulatory mechanisms, socio-economic organizations and their self-regulatory mechanisms, and even enterprise software viewed as command and control platforms for businesses. We will close with two case studies from our practice.

Biologically Inspired Hebbian Learning Function for a Simulated Robot’s Artificial Spiking Neural Network Controller

André Cyr¹, Mounir Boukadoum¹ and Pierre Poirier²

Artificial Spiking Neural Networks (ASNN) are powerful models for addressing cognitive problems, especially in the time domain. These computational tools represent a good trade-off between deep biological models of dynamic neurons and traditional artificial neural networks. Several ASSN architectures have been designed for robotic implementation and cognitive modelling in artificial life. There have also been many efforts to create diversity in their learning functions, but there exists ample room for improvement. This work presents an original Hebbian learning rule that may help gain insight into many aspects of natural neural processes with regards to the spike-timing dependent plasticity (STDP). This enhanced model of associative learning is demonstrated via a procedural and discrete-time ASNN algorithm that includes the STDP function. We present the resulting ASNN model as the brain controller of a simulated robot application. This small embodied cognitive system is validated on a 2D maze environment where contextual cues are learned online to highlight adaptive behaviors. Hence, the temporal co-occurrence of spikes is analyzed and understood not just at the synaptic level but also from the whole situated animat behavior perspective.

Keywords: robot, spike coding, Hebbian learning, adaptive behavior, artificial intelligence.

Reinforcement Sensitivity Theory and Cognitive Architectures

Karl Fua, Ian Horswill, Andrew Ortony and William Revelle

Although not a necessary distinction for purely engineering purposes, many biological models of human motivation and behavior posit a functional division between those subsystems responsible for approaching positive stimuli, and those responsible for avoiding negative stimuli. Gray and McNaughton's revised Reinforcement Sensitivity Theory (RST) cast this distinction in terms of a Behavioral Activation System (BAS) and a Fight-Flight-Freeze System (FFFS), mediated by a third, conflict resolution system -- the Behavioral Inhibition System (BIS). They argued that these are fundamental, functionally distinct systems which, among other things, have the interesting consequence of leading to a principled separation between the emotions fear and anxiety. The model has been highly influential both in personality psychology, where it provides a biologically-based explanation of the traits such as extraversion and neuroticism, and in clinical psychology wherein personality disorders such as Major Depressive Disorder (MDD) and Generalized Anxiety Disorder (GAD) can be modeled as differences in baseline sensitivities of one or more of the systems. In this paper, we present work in progress in implementing a highly simplified simulation of RST in a set of embodied virtual characters. We argue that RST provides an interesting and potentially powerful starting point for cognitive architectures for applications such as interactive entertainment, in which simulation of human-like affect and personality is important.

Taking a Mental Stance Towards Artificial Systems

David Gamez and Igor Aleksander

This paper argues that supervised cognitive growth in artifacts will be very difficult to achieve without detailed knowledge about systems’ internal states. Physical information is too low level to provide a useful understanding of a system’s behavior, and it is more pragmatically useful to take a mental stance towards an artificial system and interpret its actions in terms of mental states. This mental stance is similar to Dennett’s intentional stance, except that the ascription of beliefs and rationality in the intentional stance is replaced by the attribution of low level mental states in the mental stance. In some cases it might also be useful to take a conscious stance towards an artificial system that interprets its behavior as the outcome of a conscious decision making process. Since most artifacts lack language, automatic analysis techniques have to be used to identify the contents of their minds, and the second half of this paper suggests how some of the earlier work of Aleksander and Atlas can be applied to this area.

Measuring Rates of Human Memory Retrieval

Robert S. Gardner, Matteo Mainetti, and Giorgio A. Ascoli

A quantitative description of consciously retrievable memory is fundamental to understanding the structure of human cognition. This research utilized an automated paging procedure to measure elements of autobiographical memory (AM) and prospective memory (PM). We defined AM as recollections of personally-relevant past events, each specific to a time and a place; and PM as recollections of intentions to act within a window of time in the personal future. Fifty-three participants, recruited from an undergraduate research pool, were randomly and repeatedly paged during participant-specific predetermined time periods. At the moment a page was received, each participant evaluated the concurrence of the page with an AM or PM. If there was such a concurrence, the type of memory and corresponding duration, in seconds, were documented. We directly measured the probability of memory retrieval on a subject-by-subject basis by dividing the number of memories each individual reported by the number of their pages. Moreover, we derived the typical number of memories experienced per hour (MPH) from the probability and duration data. On average, 15 AMs and 19 PMs were experienced in one given hour. Measures of AM and PM were highly correlated across probability (r=0.602, p<0.001), and duration (r=0.912, p<0.001). Subjective estimations of the probability, duration, and number of MPH were also collected from a separate sample population. The estimated MPH was significantly lower than the corresponding value calculated from the experimental data (t=4.146, p=0.000). Altogether, the results provide a quantitative characterization of two distinct forms of consciously retrievable memory.

A Deeply-Connected Hybrid Neural-Symbolic Cognitive Architecture

Ben Goertzel

A deeply-connected hybrid neural-symbolic cognitive architecture is defined as one in which the neural-net and symbolic components interact frequently and dynamically, so that each intervenes significantly in the other's internal operations, and the two form a combined dynamical system at the time-scale of each component's individual cognitive operations. An example architecture of this nature that is currently under development is described, based on integration of the OpenCog cognitive architecture (which incorporates symbolic, evolutionary and connectionist aspects) with a hierarchical attractor neural network (HANN). In this integrated architecture, the neural and non-neural aspects each play major roles, and the depth of the interconnection is revealed for example in the facts that symbolic reasoning intervenes in the process of attractor formation within the HANN, whereas the HANN plays a major role in guiding the individual steps of logical inference and evolutionary program learning processes.

To Cognize is to Categorize Revisited: Category Theory is Where Mathematics Meets Biology

Jaime Gomez and Ricardo Sanz

This paper claims for a shift towards "the formal sciences" in the cognitive sciences. In order to explain the phenomenon of cognition, including aspects such as learning and intelligence, it is necessary to explore the concepts and methodologies offered by the formal sciences such as: dynamic system theory, category theory, control theory, graph theory, stochastic processes, probability theory and queuing theory. In particular, category theory is proposed as the most fitting tool for the building of an unified theory of cognition. This paper proposes a radically new view based in category theory is provided. A cognitive model is informally defined as a mapping between two different structures, while a structure is the set of components of a system and their relationships. Put formally in categorical terms, a model is a functor between categories that reflects the structural invariance between them. In the paper, the theory of categories is presented as the best possible framework to deal with complex system modeling -ie: biologically inspired systems that trascend and offer a much more powerful tool kit to deal with the phenomenon of cognition that other purely verbal tools like the psychological categories that Rosch or Harnad refer.

Autonomous Adaptive Brain Systems and Neuromorphic Agents

Stephen Grossberg

[abstract will appear in the proceedings]

From Functional Embodied Imagination to Episodic Memory

Owen Holland¹ and Hugo Gravato Marques²

Episodic memory is universally accepted as one of the key elements of human conscious experience, closely linked to the construction of the self, and to the sense of a continuing personal identity. Its identification by Tulving (e.g. Tulving 1984) gave rise to a range of studies by psychologists, most of which concentrated on the relationship between episodic memory and other forms of memory. However, recent work in neuropsychology has shown that episodic memory is most closely related to imagination, and the idea that in evolutionary terms it probably followed imagination, and re-used many of the structural elements supporting imagination, is now well established. This paper builds on our previous work in developing and demonstrating biologically inspired architectures for functional embodied imagination (Marques and Holland, 2009), and examines the additional mechanisms that must be added to an artificial system capable of imagination in order to give it the ability to support episodic memory. The focus throughout is on the structure and nature of the necessary architectures and component processes; this is in contrast to many existing approaches to episodic memory in artifacts, most of which have concentrated on the benefits for an agent or robot of having episodic memory, such as being able to engage in case-based reasoning. The paper also considers the relationships between episodic memory and conscious experience in the context of the model of consciousness proposed by Holland (2007).

• Holland, Owen (2007) A strongly embodied approach to machine consciousness. Journal of Consciousness Studies Special Issue on Machine Consciousness, Eds R. Chrisley, R. Clowes, and S. Torrance.
• Marques, Hugo Gravato and Holland, Owen (2009) Architectures for functional imagination. Neurocomputing, 72, 743-759
• Tulving, E. (1984). Precis of Elements of Episodic Memory. Behavioural and Brain Sciences, 7, 223 – 268.

Neural Cognitive Architectures

Christian Huyck and Emma Byrne

A neural cognitive architecture would be an architecture based on simulated neurons, that provided a set of mechanisms for all cognitive behaviour. Moreover, this would be compatible with biological neural behaviour. The development of a neural cognitive architecture is in its infancy. A proto-architecture in the form of behaving agents based on simulated neurons is described. These agents take natural language commands, view the environment, plan and act. The development of these agents has led to a series of questions that need to be addressed to advance the development of neural cognitive architectures. These questions include biological architectures including neural models and brain topology; long posed questions where progress has been made, such as the binding and symbol grounding problems; issues of emergent behaviour such as short and long-term Cell Assembly dynamics; and issues of learning such as the stability-plasticity dilemma. These questions can act as a road map for the development of neural cognitive architectures.

Neural Network Architecture for Crossmodal Activation and Perceptual Sequences

Magnus Johnsson¹, Christian Balkenius¹, and Germund Hesslow²

We present a neural network architecture with two novel features. First it uses a variant of selforganizing sensory maps that can integrate input from several sensory modalities. Second, such an associative selforganizing map can, by using recurrent associative connections while simultaneously associating its activity with the activity of another modality, be made to function as a memory of perceptual sequences. This has relevance in e.g. the modeling of expectations in one modality due to the activity occurring in another modality...

[the rest of the abstract may appear in the proceedings]

Toward Bootstrap Learning of the Foundations of Commonsense Knowledge

(invited talk)

Benjamin Kuipers

Our goal is for an autonomous learning agent to acquire the knowledge that serves as the foundations of common sense from its own experience without outside guidance. This requires the agent to (1) learn the structure of its own sensors and effectors; (2) learn a model of space around itself; (3) learn to move effectively in that space; (4) identify and describe objects, as distinct from the static environment; (5) learn and represent actions for affecting those objects, including preconditions and postconditions, and so on. We will provide examples of progress we have made, and the roadmap we envision for future research.

Biographical Sketch

Benjamin Kuipers joined the University of Michigan in January 2009 as Professor of Computer Science and Engineering. Prior to that, he held an endowed Professorship in Computer Sciences at the University of Texas at Austin. He received his B.A. from Swarthmore College, and his Ph.D. from MIT. He investigates the representation of commonsense and expert knowledge, with particular emphasis on the effective use of incomplete knowledge. His research accomplishments include developing the TOUR model of spatial knowledge in the cognitive map, the QSIM algorithm for qualitative simulation, the Algernon system for knowledge representation, and the Spatial Semantic Hierarchy model of knowledge for robot exploration and mapping. He has served as Department Chair at UT Austin, and is a Fellow of AAAI and IEEE.

Biologically Inspired Computing in CMOL CrossNets

(keynote)

Konstantin K. Likharev

CMOL is a hybrid circuit consisting of a silicon (CMOS) chip augmented with a nanowire crossbar with crosspoint devices of the “latching switch” functionality. Currently, CMOL circuits are a subject of intensive development by several academic and industrial groups, mostly in the view of the exciting prospects they provide for terabit-scale memories and reconfigurable digital logic circuits. On the other hand, the CMOL circuit is a hardware fabric uniquely suitable for the implementation of neuromorphic networks (“CrossNets”). In such networks, cell somas are realized the CMOS subsystem, crossbar nanowires play the roles of axons and dendrites, and crosspoint latching switches serve as elementary (binary-weight) synapses. We have shown that the binary character of the elementary synapses and a relatively high defect density (possible at the initial stage of CMOL technology development) do not prevent CrossNets from performing essentially all the tasks demonstrated earlier with software-implemented artificial neural networks, including associative memories, pattern classifiers, and dynamic controllers in conditions of instant and delayed reward. The significance of these results is in the very high potential areal density of CMOL CrossNets (beyond that of the mammal cerebral cortex, at similar connectivity), and the very high operation speed of these networks – e.g., intercell latency below 1 microsecond at readily manageable power dissipation below 1 W/cm². We believe that CMOL CrossNets is the first hardware which may eventually challenge the human cortex. I will discuss several examples showing how our CMOL CrossNets could be effectively used for the implementation of cognitive architectures.

An Engineering view of Cognitive Architectures

Steve Morphet

When designing a cognitive architecture, two important questions [among many] need to be answered. Specifically: (1) What components comprise a minimal cognitive substrate? (2) Which level of the biological cognitive hierarchy should be focused on to effectively develop a cognitive architecture? These questions are tightly coupled, and their answers are largely driven by the underlying implicit assumptions. For instance, is there an implication that the cognitive architecture will perform tasks that require human-level intelligence? This begs the question "what is trying to be achieved?" When a sufficiently detailed and accurate response to this question is formulated, one will be left with the set of necessary behaviors. Taking inspiration from biology and given the necessary behaviors, it is possible to delineate a minimum set of biologically inspired modules that when coupled correctly will exhibit the desired behavioral properties. It is difficult to know a priori what level of biological resolution will serve best. This is a function of the underlying task(s). But the answer will also answer the question "At what level do we as researchers need to understand how the brain works (and what it does)?" At a very low level of the resolution hierarchy, one can consider the neuron to be an atomic computation element. At a higher level, a neural cluster is a functional element, with capabilities including: arithmetic or logical operations, correlations, coordinate transformations, etc. At an even higher, the cortical column is a collection of neural clusters able to perform: focus of attention, classification, etc. When the best level of resolution is determined, there is still the question of how to emulate the lower levels still exists. Fortunately, the concept of function equivalence comes to the rescue. Specifically, functional equivalence is defined as producing the same input/output behavior regardless of underlying implementation. So, for instance, it is not necessary to implement neural clusters as neurons; rather, neural clusters implemented as a lookup table with sufficient fidelity would be not only adequate, but indistinguishable from the neural implementation. The engineering benefit: efficient software emulations can turn intractable problems into tractable ones. Additionally, the right level of resolution helps dictate the functional modules. Another important question regarding the design of a cognitive architecture: how should the modules by "wired" together? To answer this question, one can examine the role of evolution. Looking at species development and the impact of evolution on human cognitive wet-ware, one can make an argument for (1) each brain area having a fixed working function; (2) each function being implemented in overlapping neural structures. Giving these statements some thought, they should be intuitively satisfying. Unfortunately the second is contrary to state-of-the-art engineering principles. That does not necessarily make it bad; it does however present a challenge to the current thought processes in software design and reuse. The most prominent engineering implications of this insight are: (1) modularity in its classic form must be abandoned; (2) assigning computational/cognitive roles to brain areas will require cross-domain modeling; (3) cross-domain uses must be considered at design time.

Competitive Dynamics of Emotion-Cognition Interaction

Mikhail I. Rabinovich and Mehmet K. Muezzinoglu

The popular opinion that "inventing an artificial brain would be much easier than discovering the human brain" has turned out to be true. Artificial intelligence owes this success to a privilege granted to it, namely the freedom to stay detached from natural mechanisms of intelligence. Many remarkable developments in AI have enjoyed this flexibility, together with a rigorous mathematical workspace, for mimicking, and even improving mental functions. Obviously, such kind of pragmatic modeling does not (and need not) target the greatest secret of the nature, namely the origin of thoughts. The quest for this origin, however, can enrich the artificial intelligence research with invaluable tools and motivating observations. In this work, we unfold a new paradigm in the study of brain mental dynamics departing from the stable transient activity neural networks, as supported by experiments. This new approach builds upon a new intuition in contrast to traditional system analysis usually adopted in cognitive modeling. Transient dynamics offers a sound formalism of the observed qualities of brain activity, while providing a rigorous set of analysis tools. Transients have two main features: First, they are resistant to noise, and reliable even in the face of small variations in initial conditions; the sequence of states visited by the system (its trajectory), is thus structurally stable. Second, the transients are input-specific, and thus convey information about what caused them in the first place. This new dynamical view manifests a rigorous explanation of how perception, cognition, emotion, and other mental processes evolve as a sequence of activity patterns in the brain, and, most importantly, how they interfere with each other. The ideas discussed and demonstrated here lead to the creation of a quantitative theory of the human mental activity. We are sure that they can be directly deployed artificial agents as well.

Keywords: Brain dynamics; Emotion-cognition interaction; Computational modeling; Transient dynamics; Structural stability.

A Simple Oscillatory Short-Term Memory Model

(keynote)

James Reggia, Jared Sylvester, Scott Weems, and Mike Bunting

Oscillatory neural network models have been an increasing focus of activity over the last several years. These models consist of recurrent neural networks whose dynamics are characterized by persistent learned/designed rhythmic activity. Here we consider simple oscillatory memories for short-term retention of items occurring as temporal sequences. By incorporating decay as well as interference, we find that it is easy to match behavioral data from human subjects recalling temporal sequences under different situations by adjusting a single parameter in the model. These results suggest that simple oscillatory memories capture at least some key properties of human short-term memory, and might be used effectively in future cognitive architectures.

Using a Cognitive Architecture to Model the Effects of Stress and Caffeine

Frank E. Ritter¹, Sue Kase¹, Laura Cousino Klein¹, Jenette Bennett¹, and Michael Schoelles²

Cognitive architectures have found that they need to be embodied. We have added eyes and hands in the past. A next step is to examine how cognition changes based on the task and by the physiological state of the agent. We have recently concluded a study of cognition under stress. In this study, subjects performed mental arithmetic while having had one of three levels of caffeine (0, 200, 400 mg). Their performance varied based on the caffeine and appeared to vary based on their appraisal of the task as well. We have used an ACT-R model to explain what changed in these conditions. The model was fit to each individual's data using a super computer, and the resulting parameter changes help explain how cognition changes under stress and through caffeine. Our results suggesting that verbal fluency changes under stress, and that caffeine influences performance according to an inverted-U shaped curve. This work shows how fitting individuals is now possible, and how this approach can change modeling methodology.

Concepts from Data I: State-Space Partitioning with K-D Trees

Brandon Rohrer

Creating new concepts from data is a hard problem in the development of cognitive architectures, but one that must be solved for the BICA community to declare success. Three biologically-inspired concept generation algorithms are presented here that are appropriate to different levels of concept abstraction: state-space partitioning with K-D trees, temporal-sequence chunking, and context-based similarity.

Analysis of the Web User Behavior with a Psychologically Based Diffusion Model

Pablo E. Román and Juan D. Velásquez

We present a rater different approach based on a mathematical theory of psychological behavior from Usher and McClelland and the random utility model from McFadden. Those models are widely successfully tested on variety of psychological and economics studies and we pretend to use in the web user behavior modeling. The model pretend to describe the stochastic behavior of a very general class of web users consisting on the probability of choosing to follows and hyperlink or leave the site on a specific time. The new approach is similar to the data mining one, in the sense that the model’s parameter need to be fitted based on web data. The difference comes from the possibility to predict the effect of small changes on web site content and structure on the web user behavior, originating an adapted-to-human machine learning system with the possibility to have a precise way to automatic adaptation of the web site to the user. Further applicability to others human related processes are discussed and an artificial system that accurately simulate web users is presented.

• M. Usher and J. McClelland. The time course of perceptual choice: The leaky, competing accumulator model. Psychological Review, 108:550-592, 2001.
• McFadden, D., ls Conditional Logit Analysis of Qualitative Choice Behavior, in Zarembka (ed.), Frontiers in Econometrics, Academic Press, 1973.

From Brain Models to Cognitive Architectures

Ricardo Sanz and Jaime Gomez and Carlos Hernandez and Adolfo Hernando

The construction of biologically inspired cognitive systems is based in the extraction of functional patterns from biological systems. In particular, functional brain models constitute the basis for many ongoing cognitive systems construction projects. Brain models are the core intermediate asset in this scientific and engineering process. The implementation of brain models as computational systems running atop simulators or physical robots serve the two proposes of i) validating the functional competences captured in the model and ii) serving as blueprints for artificial cognitive systems. However, the validity and research method of this whole process is compromised by the fact that the transition from models to implementations is done by handcrafting a bunch of code. The pretended isomorphism between models and implementations is false in general. Models are tested by implementation but this test is only as good as the isomorphism holds. In this paper we will analise this problem and suggest a solution based on the use of rigourous model-to-code transformation engines.

The New Schema Formalism of Cognitive Constructor: A Step Beyond the Object-Oriented Paradigm

Alexei V. Samsonovich

The paper will provide an abridged description of the formalism of schemas developed as a part of the Cognitive Constructor architecture, the successor of GMU-BICA (the familiar term “schema” is used to refer to a new concept because of the lack of a better term, and may be a temporary solution). “Schemas” are contrasted to traditional objects, schemas, classes, frames, etc. The agenda of this study is (a) to compare principles of the two formalisms: OOP and “schemas” and to observe that they are not easily reducible to each other; (b) to illustrate by examples the advantage of the new formalism over the traditional OOP, and (c) to put “schemas” in the context of Cognitive Constructor, connecting them to the formalism of mental states (Samsonovich et al., IJMC 1.1.111-130.2009). Cognitive Constructor is a self-aware cognitive architecture inspired by the human mind.

The MGLAIR Cognitive Architecture

(keynote)

Stuart C. Shapiro

MGLAIR (Modal Grounded Layered Architecture with Integrated Reasoning) is a multi-layered cognitive architecture for embodied agents operating in real, virtual, or simulated environments containing other agents. The lowest layer, the Sensori-Actuator Layer (SAL), contains the controllers of the agent's sensors and effectors. The highest layer, the Knowledge Layer (KL), contains the agent's beliefs, including: semantic and episodic memory; quantified and conditional beliefs used for reasoning; plans for carrying out complex acts and for achieving goals; preconditions and effects of acts; policies about when, and under what circumstances, acts should be performed; self-knowledge; and metaknowledge. The KL is the layer in which conscious reasoning, planning, and act execution is performed. Between the KL and the SAL is the perceptuo-motor layer (PML), which, itself is divided into three sublayers. The lowest of these, the PMLc, directly abstracts the sensors and effectors into the basic behavioral repertoire of the agent's body. The highest, the PMLa, contains: the subconscious implementation of the cognitively primitive actions of the KL; the structures used for the perception of objects and properties in the environment; various registers for providing the agent's sense of situatedness in the environment, such as its sense of "I", "You", "Now", and the actions it is currently engaged in; and procedures for natural language comprehension and generation. The middle PML sublayer, the PMLb handles translation and communication between the PMLa and the PMLc.

The KL constitutes the mind of the agent; the PML and SAL, its body. However, the KL and PMLa layers are independent of the implementation of the agent's body, and can be connected, without change, to a hardware robot, or to a variety of software-simulated robots or avatars.

The sensors and effectors can operate simultaneously. To take advantage of this, MGLAIR is organized into modalities. A modality represents a limited resource---a PMLc-level behavior that is limited in what it can do at once, but is independent of the behaviors of other modalities. Each modality runs in a separate thread, and uses its own channel for communication between the PMLb and PMLc layers. Modalities that have been implemented in various MGLAIR agents include speach, hearing, navigation, and vision.

Experiments on the Acquisition of Linguistic Competence in Communicating Propositional Logic Sentences

Josefina Sierra

We describe some experiments which simulate a grounded approach to language acquisition in a population of autonomous agents without prior linguistic knowledge. The idea is to let the agents acquire at the same time a conceptualisation of their environment and a shared language (lexicon and grammar) which allows them to express facts about their environment in a way that could be understood by other agents in the population. The approach used to simulate the conceptualisation and the language acquisition processes in each individual agent is based on general purpose cognitive capacities, such as visual perception, categorisation, discrimination, evaluation, invention, adoption and induction. The emergence of a shared language in the population, and therefore the acquisition of a common set of linguistic conventions by the individual agents, results from a process of self-organisation of a particular type of linguistic interaction, known as a language game, which takes place among the agents in the population.

The experiments show that at the end of a simulation run the agents build different conceptualisations and different grammars. However, these conceptualisations and grammars are compatible enough to guarantee the unambiguous communication of propositional logic formulas.

We also show that the categorisers of the perceptual and logical categories built during the conceptualisation and language acquisition processes can be used for some forms of common sense reasoning, such as determining whether a sentence is a tautology, a contradiction, a common sense axiom or a merely satisfiable formula, in a very restricted domain. Although this form of intuitive reasoning requires the agents to be conscious of the fact that they use certain categorisers and of the behaviour of such categorisers.

Time Production and Representation in a Conceptual and Computational Cognitive Model

Javier Snaider, Ryan McCall, and Stan Franklin

Time perception is unlike other modes of perception such as motion, color or sound. We can see images, listen to sounds, or touch objects. Our senses enable us to perceive events of the real world. But time is different. We don’t have a sense for time. The concept of time is integral to the cognitive process. Here we will argue that, instead of asking “How can time be perceived?”, we should ask “How is a sense of time produced by a cognitive system?”. In other words, we contend that time is something that the cognitive process constructs.

An autonomous agent can be defined as “A system embedded in, and part of, an environment that senses its environment acts on it over time in pursuit of its own agenda so that its actions affect its future sensing”. We argue that the perception of time, its representation, interpretation and manipulation, are crucial abilities for many autonomous agents.

We present a model for time perception that concentrates on succession and duration, two important cognitive aspects of time. It generates these concepts and others, such as continuity, immediate present duration, and lengths of time. We also address event hierarchy and expectations as subjects intimately related with time and temporal perspective.

We illustrate these ideas using the LIDA cognitive model; however, this model for time is general enough to be implemented in any sufficiently comprehensive cognitive architecture.

Dopamine, Learning, and Production Rules: The Basal Ganglia and the Flexible Control of Information Transfer in the Brain

Andrea Stocco, Christian Lebiere, and John R. Anderson

One of the open issues in developing large-scale computational models of the brain is how the transfer of information between communicating cortical regions is controlled. A solution to this problem ultimately provides insight into broader questions such as the nature of cognitive control and the organization of intelligent behavior.

Here, we present a model exploring the hypothesis that the basal ganglia implement such a conditional information routing system. The basal ganglia are a set of subcortical nuclei that play a central role in cognition. Like a switchboard, the model basal ganglia direct the communication between cortical regions by alerting the destination regions of the presence of important signals coming from the source regions. This way, they can impose serial control on the massive parallel communication between cortical areas.

We suggest that such a mechanism provides an account for several cognitive functions of the basal ganglia. The model also incorporates a possible mechanism by which subsequent transfers of information control the release of dopamine. This signal is used to produce novel stimulus-response associations by internalizing the representation being transferred in the striatum. Such representations remain available even in absence of cortical stimulation, providing a shortcut to automatic behaviors.

We discuss how this neural circuit can be seen as a biological implementation of a production system. This comparison highlights the basal ganglia as bridge between computational models of small-size brain circuits and high-level characterizations of complex cognition, such as cognitive architectures. Finally, a series of simulations is presented, which illustrate how the model can perform simple stimulus-response tasks, develop automatic behaviors, and provide an account for the cognitive and motor symptoms of basal ganglia pathologies.

[Title TBA]

(invited joint session talk)

Hans-Georg Stork

[abstract not available at this time]

A Brain-Inspired Cognitive Architecture for Reinforcement Learning

Tyler Streeter^1,2 and James Oliver¹

We present a novel cognitive architecture, Sapience, inspired by the high-level organization of the mammalian brain. This architecture include five components based on abstract functional representations of the brain’s sensorimotor cortex, hippocampus, basal ganglia, cerebellum, and prefrontal cortex regions. Each component provides a unique computational benefit to the system. We intend it to be applicable generally to all kinds of applications involving autonomous high-dimensional real-time learning and control, including video games and robotics. Crucially, it is designed to be practical: fully implementable in software, scalable in performance based on available computing hardware, and general enough to be applied to a wide variety of control problems.

The objective of our system is to learn complex motor control tasks defined as reinforcement learning problems. The system’s behavior is shaped by two sources of reinforcements: 1) external rewards for achieving goals, defined by the programmer, and 2) internal rewards for improved understanding of the world, aka “curiosity rewards”. This curiosity drive is based on recent theoretical work in artificial curiosity which provides a powerful model of autonomous self-development.

Our contribution is a concrete, implementable, brain-inspired cognitive architecture which integrates several key features: high-dimensional sensory and motor arrays; learned topographic maps for arbitrary sensor arrangements; a probabilistic symbolic world model (hierarchical Bayesian network) utilizing robust, adaptable kernel mixture models for the conditional probability distributions; sequential prediction learning; reinforcement learning of context-dependent value and action selection; curiosity-driven world model improvements; supervised learning of parallel action selection; and a temporary working memory with gated read/write actions, enabling reinforcement learning of arbitrary programs.

HELEN: Hierarchical Event-Language Network for Knowledge Representation

Robert Swaine

A biologically constraint model is presented: HELEN Hierarchical Event-Language Network. The model has the goal of representing certain portions of human situational experience to model natural language and reason situationally. Local brain regions and pathways are structured in the hierarchies mimicking the sensory streams from the brainstem, throughout the cortex, to the cingulate regions. From this, segmented situations represented by the model are grouped and sequenced into an event-language that can index patterns of situations. This event language is used to model natural language innately as a story understander. Emotions, procedural actions, spatial relationships, are three of the four key domains used in the model.

From Constructionist to Constructivist AI

Kristinn R. Thórisson

The development of artificial intelligence systems has been largely one of manual labor. Following -- and sometimes slightly ahead of -- standard software methodologies, the AI developer is essentially in the role of "construction worker". This Constructionist approach to AI is based on the long tradition of divide-and-conquer in science: divide the problem up small enough that a student or a small team of researchers can find a solution within a few years. To no surprise, the result is a diverse set of isolated solutions to relatively small problems. Some success of putting the pieces back together in other fields of science, most notably physics, has spurred optimism, in spite of such progress being slow, extending over centuries and millennia. The way "the AI problem" has been divided up to date has been driven by concerns for potential application areas, not scientific or theoretical insight. The result is an unavoidable fragmentation of the research community and a set of grossly incompatible approaches to the various dismembered (sub)parts of the problem. Although today's AI systems clearly are useful in some ways, the Constructionist approach to AI has shown itself to result in systems with limited domain application and severe performance brittleness. Ever-larger "idiot savant" systems are being built that show little or no hope of growing beyond their still highly-limited operating contexts, falling far of what could reasonably pass for general intelligence. Standard software development methods do not scale well, as anyone knows who has tried to coordinate more than 5 or 6 executables running in parallel. In AI, "putting Humpty-Dumpty back together again" is not simply a huge undertaking, it is practically and theoretically intractable. Yet going beyond current AI systems requires integrating significantly more complex processes than attempted to date, in vastly greater numbers. Clearly this calls for a radically different approach. I argue that the only way to address the challenge is replacing hand-crafting and top-down architectural design with self-generated code and self-organizing architectures. I call this Constructivist AI, in reference to auto-constructive systems, i.e. those that partly or fully construct themselves. It stands to reason that the methodologies employed for Constructivist AI will be very different from today's software development methods. In this talk we will examine some of the implications of the impending Constructivist AI paradigm shift.

Dr. Kristinn R. Thórisson is Associate Professor at the School of Computer Science, Reykjavik University, and founding director of the Center for Analysis and Design of Intelligent Agents (http://cadia.ru.is). Kris has been researching AI in academia and industry for two decades. His work has been focused on large-scale architecture for cognitive systems, from pioneering work on integrated multimodal perception, knowledge and action control at MIT, to directing the design and implementation of systems with spoken and written conversational skills in New York, to the development of large virtual worlds at LEGO. He has consulted on technology and business for various companies including British Telecom and NASA and he is co-founder of Radar Networks Inc. in California. He is Coordinator of the multi-year HUMANOBS research project, funded by the European Union, which aims at building learning, self-modifying AI architectures. Kris holds a Ph.D. in Media Arts and Sciences from the MIT Media Laboratory.

http://www.ru.is/faculty/thorisson

Emotions: A Bridge Between Nature and Society?

Rodrigo Ventura

The field of Artificial Intelligence has for a long time neglected, the role of emotions in human cognition, with few but notable exceptions [Simon76]. This has been motivated in part by the assumption that the emulation of human rationality by a machine is sufficient for attaining general human-level intelligence. This paper reviews neuroscientific results that have been showing empirical evidence, consistently for over a decade, sustaining that the emotion mechanisms in the brain play a fundamental role on its decision making processes, as well as on cognitive regulation. Moreover, this role takes place regardless whether the subject is aware of any emotion. These mechanisms are particularly important in social contexts. Lesions in the pathways supporting these mechanisms provoke serious impairments on social behavior. For instance, subjects with lesions in the pathways between the orbitofrontal cortex and the amygdala are no longer able to sustain an healthy social live, despite their intact intellectual capabilities [Damasio94]. Strikingly, these patients are even able to verbally describe what would be the proper social behavior, although are unable to follow it. One important mechanism in social contexts is empathy, fundamental for proper social relations. It has been proposed that empathy is founded on mechanisms analogous to the mirror neurons [Carr03,Jackson05]. Following the broad literature in neuroscience on emotions and cognition function, many computational models of emotions, addressing its role in cognition, have been developed.

• [Simon76] Herbert A. Simon, "Motivational and Emotional Controls of Cognition", Psychological Review 74(1), 1967.
• [Damasio94] Antonio R. Damasio, "Descartes' Error: Emotion, Reason and the Human Brain", Picador, 1994.
• [Carr03] Laurie Carr, Marco Iacoboni, Marie-Charlotte Dubeau, John C. Mazziotta, Gian Luigi Lenzi, "Neural mechanisms of empathy in humans: A relay from neural systems for imitation to limbic areas", PNAS 100(9), 2003.
• [Jackson05] Philip L. Jackson, Andrew N. Meltzoff, Jean Decety, "How do we perceive the pain of others? A window into the neural processes involved in empathy", NeuroImage 24, 2005.

Insufficient Knowledge and Resources - A Biological Constraint and Its Functional Implications

Pei Wang

Insufficient knowledge and resources is not only a biological constraint on human and animal intelligence, but also has important functional implications for artificial intelligence (AI) systems. Traditional theories dominating AI research typically assume some kind of sufficiency of knowledge and resources, so cannot solve many problems in the field. AI needs new theories obeying this constraint, which cannot be obtained by minor revisions and extensions of the traditional theories. The practice of NARS, an AGI project, shows that such new theories are feasible and promising in providing a new theoretical foundation for AI.

A Complete Ethical System (and Supporting Architecture) for Intelligent Machines

Mark Waser

As machines become more intelligent and take on additional responsibilities, their decision-making capabilities must be informed and constrained by a common sense moral/ethical structure for everyone’s safety and well-being. Unfortunately, “common sense” is not so common and human society itself frequently disagrees on what is ethical and what is not with many individuals arguing that ethics is subjective and most believing that no reasonably simple foundation exists for the determination of the correctness or morality of any given action. It is our contention, however, that these apparent difficulties are solely due to the fact that, while human beings have evolved to have a moral/ethical system, that system is still incomplete and frequently overridden by an inefficient and unintelligent self-centered shortsightedness that has already become a clear evolutionary mismatch. We propose to solve these problems by a) showing that morality is actually objective and derivable from first principles; b) presenting a coherent, integrated, platonic ethical system with no internal inconsistencies that flows naturally from a single high-level logically-derived Kantian imperative to low-level reflexive "rules of thumb" that match current human sensibilities; and c) suggesting a biologically-inspired architecture which supports and enforces this system which can be relatively easily implemented.