Search

A categorical and Łukasiewicz-Topos framework for Algebraic Logic models of nonlinear dynamics in complex functional systems such as Neural Networks, Cell Genome and Interactome Networks is introduced. Łukasiewicz Algebraic Logic models of both neural and genetic networks and signaling pathways in cells are formulated in terms of nonlinear dynamic systems with n-state components that allow for the generalization of previous logical models of both genetic activities and neural networks. An algebraic formulation of variable 'next-state functions' is extended to a Łukasiewicz Topos with an n-valued Łukasiewicz Algebraic Logic subobject classifier description that represents non-random and nonlinear network activities as well as their transformations in developmental processes and carcinogenesis.

Many existing independent component analysis algorithms include a preprocessing stage where the inputs are sphered. This amounts to normalising the data such that all correlations between the variables are removed. In this work, I show that sphering allows very weak contextual modulation to steer the development of meaningful features. Context-biased competition has been proposed as a model of covert attention and I propose that sphering-like normalisation also allows weaker top-down bias to guide attention.

A new algorithmic framework called denoising source separation (DSS) is introduced. The main benefit of this framework is that it allows for easy development of new source separation algorithms which are optimised for specific problems. In this framework, source separation algorithms are constucted around denoising procedures. The resulting algorithms can range from almost blind to highly specialised source separation algorithms. Both simple linear and more complex nonlinear or adaptive denoising schemes are considered. Some existing independent component analysis algorithms are reinterpreted within DSS framework and new, robust blind source separation algorithms are suggested. Although DSS algorithms need not be explicitly based on objective functions, there is often an implicit objective function that is optimised. The exact relation between the denoising procedure and the objective function is derived and a useful approximation of the objective function is presented. In the experimental section, various DSS schemes are applied extensively to artificial data, to real magnetoencephalograms and to simulated CDMA mobile network signals. Finally, various extensions to the proposed DSS algorithms are considered. These include nonlinear observation mappings, hierarchical models and overcomplete, nonorthogonal feature spaces. With these extensions, DSS appears to have relevance to many existing models of neural information processing.

A new algorithmic framework called denoising source separation (DSS) is introduced. The main benefit of this framework is that it allows for easy development of new source separation algorithms which are optimised for specific problems. In this framework, source separation algorithms are constucted around denoising procedures. The resulting algorithms can range from almost blind to highly specialised source separation algorithms. Both simple linear and more complex nonlinear or adaptive denoising schemes are considered. Some existing independent component analysis algorithms are reinterpreted within DSS framework and new, robust blind source separation algorithms are suggested. Although DSS algorithms need not be explicitly based on objective functions, there is often an implicit objective function that is optimised. The exact relation between the denoising procedure and the objective function is derived and a useful approximation of the objective function is presented. In the experimental section, various DSS schemes are applied extensively to artificial data, to real magnetoencephalograms and to simulated CDMA mobile network signals. Finally, various extensions to the proposed DSS algorithms are considered. These include nonlinear observation mappings, hierarchical models and overcomplete, nonorthogonal feature spaces. With these extensions, DSS appears to have relevance to many existing models of neural information processing.

The dentate gyrus is part of the hippocampal memory system and special in that it generates new neurons throughout life. Here we discuss the question of what the functional role of these new neurons might be. Our hypothesis is that they help the dentate gyrus to avoid the problem of catastrophic interference when adapting to new environments. We assume that old neurons are rather stable and preserve an optimal encoding learned for known environments while new neurons are plastic to adapt to those features that are qualitatively new in a new environment. A simple network simulation demonstrates that adding new plastic neurons is indeed a successful strategy for adaptation without catastrophic interference.

What type of artificial systems will claim to be conscious and will claim to experience qualia? The ability to comment upon physical states of a brain-like dynamical system coupled with its environment seems to be sufficient to make claims. The flow of internal states in such system, guided and limited by associative memory, is similar to the stream of consciousness. Minimal requirements for an artificial system that will claim to be conscious were given in form of specific architecture named articon. Nonverbal discrimination of the working memory states of the articon gives it the ability to experience different qualities of internal states. Analysis of the inner state flows of such a system during typical behavioral process shows that qualia are inseparable from perception and action. The role of consciousness in learning of skills, when conscious information processing is replaced by subconscious, is elucidated. Arguments confirming that phenomenal experience is a result of cognitive processes are presented. Possible philosophical objections based on the Chinese room and other arguments are discussed, but they are insufficient to refute claims articon’s claims. Conditions for genuine understanding that go beyond the Turing test are presented. Articons may fulfill such conditions and in principle the structure of their experiences may be arbitrarily close to human.

We study a modification of the Hoppensteadt-Izhikevich canonical model for networks of class I neurons, in which the 'pulse' emitted by a neuron is smooth rather than a delta-function. We prove two types of results about synchronization and desynchronization of such networks, the first type pertaining to 'pulse' functions which are symmetric, and the other type in the regime in which each neuron is connected to many other neurons.

We study a modification of the Hoppensteadt-Izhikevich canonical model for networks of class I neurons, in which the 'pulse' emitted by a neuron is smooth rather than a delta-function. We prove two types of results about synchronization and desynchronization of such networks, the first type pertaining to 'pulse' functions which are symmetric, and the other type in the regime in which each neuron is connected to many other neurons.

We examine a simple biologically-motivated neural network, the three-layer version of the Chialvo-Bak `minibrain' [Neurosci. 90 (1999) 1137], and present numerical results which indicate that a non-equilibrium phase transition between ordered and disordered phases occurs subject to the tuning of a control parameter. Scale-free behaviour is observed at the critical point. Notably, the transition here is due solely to network geometry and not any noise factor. The phase of the network is thus a design parameter which can be tuned. The phases are determined by differing levels of interference between active paths in the network and the consequent accidental destruction of good paths.

Human cognition is unique in the way in which it relies on combinatorial (or compositional) structures. Language provides ample evidence for the existence of combinatorial structures, but they can also be found in visual cognition. To understand the neural basis of human cognition, it is therefore essential to understand how combinatorial structures can be instantiated in neural terms. In his recent book on the foundations of language, Jackendoff described four fundamental problems for a neural instantiation of combinatorial structures: the massiveness of the binding problem, the problem of 2, the problem of variables and the transformation of combinatorial structures from working memory to long-term memory. This paper aims to show that these problems can be solved by means of neural ‘blackboard’ architectures. For this purpose, a neural blackboard architecture for sentence structure is presented. In this architecture, neural structures that encode for words are temporarily bound in a manner that preserves the structure of the sentence. It is shown that the architecture solves the four problems presented by Jackendoff. The ability of the architecture to instantiate sentence structures is illustrated with examples of sentence complexity observed in human language performance. Similarities exist between the architecture for sentence structure and blackboard architectures for combinatorial structures in visual cognition, derived from the structure of the visual cortex. These architectures are briefly discussed, together with an example of a combinatorial structure in which the blackboard architectures for language and vision are combined. In this way, the architecture for language is grounded in perception.

It is argued that the key to understanding the brain is to view it as a device making extensive use of methodologies developed in computer programming, the idea of compiling source code written in a high-level language providing a mechanism for conceptually linking the two domains. Following the argument through, one arrives at a clarification of what the nervous system in its complexity is all about; it consists of a collection of devices for implementing specific kinds of competence, in ways in principle indicated in detail by application of the object-oriented programming paradigm to the various kinds of processes featuring in cognitive life.

We investigate the behavioral patterns of a population of agents, each controlled by a simple biologically motivated neural network model, when they are set in competition against each other in the Minority Model of Challet and Zhang. We explore the effects of changing agent characteristics, demonstrating that crowding behavior takes place among agents of similar memory, and show how this allows unique `rogue' agents with higher memory values to take advantage of a majority population. We also show that agents' analytic capability is largely determined by the size of the intermediary layer of neurons. In the context of these results, we discuss the general nature of natural and artificial intelligence systems, and suggest intelligence only exists in the context of the surrounding environment (embodiment). Source code for the programs used can be found at http://neuro.webdrake.net/.

To give an adequate explanation of cognition and perform certain practical tasks connectionist systems must be able to extrapolate. This work has explored the relationship between input representation and extrapolation, using simulations of multi-layer perceptrons trained to model the identity function. It has been discovered that representation has a marked effect on extrapolation.

Technology of data visualization and data modeling is suggested. The basic of the technology is original idea of elastic net and methods of its construction and application. A short review of relevant methods has been made. The methods proposed are illustrated by applying them to the real economical, sociological and biological datasets and to some model data distributions. The basic of the technology is original idea of elastic net - regular point approximation of some manifold that is put into the multidimensional space and has in a certain sense minimal energy. This manifold is an analogue of principal surface and serves as non-linear screen on what multidimensional data are projected. Remarkable feature of the technology is its ability to work with and to fill gaps in data tables. Gaps are unknown or unreliable values of some features. It gives a possibility to predict plausibly values of unknown features by values of other ones. So it provides technology of constructing different prognosis systems and non-linear regressions. The technology can be used by specialists in different fields. There are several examples of applying the method presented in the end of this paper.

Technology of data visualization and data modeling is suggested. The basic of the technology is original idea of elastic net and methods of its construction and application. A short review of relevant methods has been made. The methods proposed are illustrated by applying them to the real economical, sociological and biological datasets and to some model data distributions. The basic of the technology is original idea of elastic net - regular point approximation of some manifold that is put into the multidimensional space and has in a certain sense minimal energy. This manifold is an analogue of principal surface and serves as non-linear screen on what multidimensional data are projected. Remarkable feature of the technology is its ability to work with and to fill gaps in data tables. Gaps are unknown or unreliable values of some features. It gives a possibility to predict plausibly values of unknown features by values of other ones. So it provides technology of constructing different prognosis systems and non-linear regressions. The technology can be used by specialists in different fields. There are several examples of applying the method presented in the end of this paper.

An associative neural network (ASNN) is a combination of an ensemble of the feed-forward neural networks and the K-nearest neighbor technique. The introduced network uses correlation between ensemble responses as a measure of distance amid the analyzed cases for the nearest neighbor technique and provides an improved prediction by the bias correction of the neural network ensemble. An associative neural network has a memory that can coincide with the training set. If new data become available, the network further improves its predicting ability and can often provide a reasonable approximation of the unknown function without a need to retrain the neural network ensemble.

In this methodological work I explore the possibility of explicitly modelling expectations conditioning the R&D decisions of firms. In order to isolate this problem from the controversies of cognitive science, I propose a black box strategy through the concept of “internal model”. The last part of the article uses artificial neural networks to model the expectations of firms in a model of industry dynamics based on Nelson & Winter (1982).

What genotypic features explain the evolvability of organisms that have to accomplish many different tasks? The genotype of behaviorally complex organisms may be more likely to encode modular neural architectures because neural modules dedicated to distinct tasks avoid neural interference, i.e., the arrival of conflicting messages for changing the value of connection weights during learning. However, if the connection weights for the various modules are genetically inherited, this raises the problem of genetic linkage: favorable mutations may fall on one portion of the genotype encoding one neural module and unfavorable mutations on another portion encoding another module. We show that this can prevent the genotype from reaching an adaptive optimum. This effect is different from other linkage effects described in the literature and we argue that it represents a new class of genetic constraints. Using simulations we show that sexual reproduction can alleviate the problem of genetic linkage by recombining separate modules all of which incorporate either favorable or unfavorable mutations. We speculate that this effect may contribute to the taxonomic prevalence of sexual reproduction among higher organisms. In addition to sexual recombination, the problem of genetic linkage for behaviorally complex organisms may be mitigated by entrusting evolution with the task of finding appropriate modular architectures and learning with the task of finding the appropriate connection weights for these architectures.

Natural autonomous systems, such as biological, neural, social or cultural systems, are open, self-organized systems with a more or less large hierarchy of interacting complexity levels; they are able to memorize their experiences and to adapt to various conditions through a change of behavior. These last fifteen years, the Authors have developed a mathematical model for these systems, based on Category Theory. The aim of the paper is to give an overview of this model, called Memory Evolutive Systems.

The schemes for compositional distributed representations include those allowing on-the-fly construction of fixed dimensionality codevectors to encode structures of various complexity. Similarity of such codevectors takes into account both structural and semantic similarity of represented structures. In this paper we provide a comparative description of sparse binary distributed representation developed in the frames of the Associative-Projective Neural Network architecture and more well-known Holographic Reduced Representations of Plate and Binary Spatter Codes of Kanerva. The key procedure in Associative-Projective Neural Networks is Context-Dependent Thinning which binds codevectors and maintains their sparseness. The codevectors are stored in structured memory array which can be realized as distributed auto-associative memory. Examples of distributed representation of structured data are given. Fast estimation of similarity of analogical episodes by the overlap of their codevectors is used in modeling of analogical reasoning for retrieval of analogs from memory and for analogical mapping.

The issue of meaningful yet unexpressed background - to language, to our experiences of the body - is one whose exploration is still in its infancy. There are various aspects of "invisible," implicit, or background experiences which have been investigated from the viewpoints of phenomenology, cognitive psychology, and linguistics. I will claim that James, as explicated by Gurwitsch and others, has analyzed the phenomenon of fringes in such a way as to provide a structural framework from which to investigate and better understand those ideas or concepts that are unexpressed, particularly those experienced in the sense of being sought-after. I will consider Johnson’s conception of the image-schematic gestalt (ISG) as a way of bridging the descriptive gap between phenomenology and cognitive psychology. Starting from an analysis of the fringes, I will turn to a consideration of the of tip-of-tongue (TOT) state, as a kind of feeling-of-knowing (FOK) state, from a variety of approaches, focusing mainly on cognitive psychology and phenomenology. I will then integrate a phenomenological analysis of these experiences, from the James/Gurwitsch structural viewpoint, with a cognitive/phenomenological analysis in terms of ISGs; and further integrate that with a cognitive/functional analysis of consciousness. I will employ this synthesis of three viewpoints to explore the thesis that the TOT state and similar experiences may relate to the gestalt nature of schemas as well as to particular cues, and may thus be experienced as an aspect of the continuum to the general background to all our conscious experiences.

Memory Evolutive Systems (MES) represent a mathematical model, based on Category Theory, to study natural open autonomous systems such as biological, neural or social systems. It has been progressively developed by the authors in a series of papers since 1986. In this model the dynamics is modulated by the competitive interactions between a net of internal more or less complex organs of regulation, called CoRegulators (CR), with a differential access to a central hierarchical Memory. This article attempts to model the notions of Semantics and Consciousness in such a MES A Semantics will emerge through the detection of specific invariances by the CRs that leads to classify objects according to their main attributes, and record the invariance classes. The model explains how it relies on a hierarchical 2 steps process: first a pragmatically 'acted' classification at the level of specific attributes (such as colors), then this classification is 'reflected' and analyzed at a higher level, and a new formal unit, called a 'concept', is formed to represent the invariance class (e.g., the color 'blue'). The introduction of more and more abstract concepts gives more flexibility to the comportment. It is essential for the development of some kind of 'consciousness'. A 'conscious' CR is characterized by the capacity to respond to a new event or to a fracture by an increase in awareness, which permits: (i) to extend its actual 'landscape' (formed by the information it can gather) retrospectively to past lower levels; (ii) to operate an abduction process in this extended landscape to find possible causes of the fracture; (iii) and finally to planify a strategy for several steps ahead, through the formation of internal 'virtual' landscapes in which strategies can be tried without energy costs. Thus consciousness would amount to an internalization of Semantics and Time, giving a selective advantage. In the second Part of the paper, a MES modeling a neural system is explicitly described and it is shown how the various processes described above are in agreement with present neurophysiological knowledge. Finally the general ideas are illustrated on a concrete example.

We review the concepts of field computation, a model of computation that processes information represented as spatially continuous arrangements of continuous data. We show that many processes in the brain are described usefully as field computation. Throughout we stress the connections between field computation and quantum mechanics, especially including the important role of information fields, which represent by virtue of their form rather than their magnitude. We also show that field computation permits simultaneous nonlinear computation in linear superposition.

Distributed representations were often criticized as inappropriate for encoding of data with a complex structure. However Plate's Holographic Reduced Representations and Kanerva's Binary Spatter Codes are recent schemes that allow on-the-fly encoding of nested compositional structures by real-valued or dense binary vectors of fixed dimensionality. In this paper we consider procedures of the Context-Dependent Thinning which were developed for representation of complex hierarchical items in the architecture of Associative-Projective Neural Networks. These procedures provide binding of items represented by sparse binary codevectors (with low probability of 1s). Such an encoding is biologically plausible and allows to reach high information capacity of distributed associative memory where the codevectors may be stored. In distinction to known binding procedures, Context-Dependent Thinning allows to support the same low density (or sparseness) of the bound codevector for varied number of constituent codevectors. Besides, a bound codevector is not only similar to another one with similar constituent codevectors (as in other schemes), but it is also similar to the constituent codevectors themselves. This allows to estimate a structure similarity just by the overlap of codevectors, without the retrieval of the constituent codevectors. This also allows an easy retrieval of the constituent codevectors. Examples of algorithmic and neural network implementations of the thinning procedures are considered. We also present representation examples of various types of nested structured data (propositions using role-filler and predicate-arguments representation, trees, directed acyclic graphs) using sparse codevectors of fixed dimension. Such representations may provide a fruitful alternative to the symbolic representations of traditional AI, as well as to the localist and microfeature-based connectionist representations.

What is it that makes a face beautiful? Average faces obtained by photographic (Galton 1878) or digital (Langlois & Roggman 1990) blending are judged attractive but not optimally attractive (Alley & Cunningham 1991) --- digital exaggerations of deviations from average face blends can lead to higher attractiveness ratings (Perrett, May, & Yoshikawa 1994). My novel approach to face design does not involve blending at all. Instead, the image of a female face with high ratings is composed from a fractal geometry based on rotated squares and powers of two. The corresponding geometric rules are more specific than those previously used by artists such as Leonardo and Duerer. They yield a short algorithmic description of all facial characteristics, many of which are compactly encodable with the help of simple feature detectors similar to those found in mammalian brains. This suggests that a face's beauty correlates with simplicity relative to the subjective observer's way of encoding it.

This paper is about how to give human-like powers to complete agents. For this the most important design choice concerns the overall architecture. Questions regarding detailed mechanisms, forms of representations, inference capabilities, knowledge etc. are best addressed in the context of a global architecture in which different design decisions need to be linked. Such a design would assemble various kinds of functionality into a complete coherent working system, in which there are many concurrent, partly independent, partly mutually supportive, partly potentially incompatible processes, addressing a multitude of issues on different time scales, including asynchronous, concurrent, motive generators. Designing human like agents is part of the more general problem of understanding design space, niche space and their interrelations, for, in the abstract, there is no one optimal design, as biological diversity on earth shows.

The relation between mental processes and brain activity is studied from the point of view of the theory of the organism-environment system. It is argued that the systemic point of view leads to a new kind of definition of the primary tasks of neurophysiology and to a new understanding of the traditional neurophysiological concepts. Neurophysiology is restored to its place as a part of biology: its task is the study of neurons as living units, not as computer chips. Neurons are living units which are organised as metabolic systems in connection with other neurons; they are not units which would carry out some psychological functions or maintain states which are typical only of the whole organism-environment system. Psychological processes, on the other hand, are processes always comprising the whole organism-environment system.

There is now a huge amount of interest in consciousness among scientists as well as philosophers, yet there is so much confusion and ambiguity in the claims and counter-claims that it is hard to tell whether any progress is being made. This ``position paper'' suggests that we can make progress by temporarily putting to one side questions about what consciousness is or which animals or machines have it or how it evolved. Instead we should focus on questions about the sorts of architectures that are possible for behaving systems and ask what sorts of capabilities, states and processes, might be supported by different sorts of architectures. We can then ask which organisms and machines have which sorts of architectures. This combines the standpoint of philosopher, biologist and engineer. If we can find a general theory of the variety of possible architectures (a characterisation of ``design space'') and the variety of environments, tasks and roles to which such architectures are well suited (a characterisation of ``niche space'') we may be able to use such a theory as a basis for formulating new more precisely defined concepts with which to articulate less ambiguous questions about the space of possible minds. For instance our initially ill-defined concept (``consciousness'') might split into a collection of more precisely defined concepts which can be used to ask unambiguous questions with definite answers. As a first step this paper explores a collection of conjectures regarding architectures and their evolution. In particular we explore architectures involving a combination of coexisting architectural levels including: (a) reactive mechanisms which evolved very early, (b) deliberative mechanisms which evolved later in response to pressures on information processing resources and (c) meta-management mechanisms that can explicitly inspect evaluate and modify some of the contents of various internal information structures. It is conjectured that in response to the needs of these layers, perceptual and action subsystems also developed layers, and also that an ``alarm'' system which initially existed only within the reactive layer may have become increasingly sophisticated and extensive as its inputs and outputs were linked to the newer layers. Processes involving the meta-management layer in the architecture could explain the origin of the notion of ``qualia''. Processes involving the ``alarm'' mechanism and mechanisms concerned with resource limits in the second and third layers gives us an explanation of three main forms of emotion, helping to account for some of the ambiguities which have bedevilled the study of emotion. Further theoretical and practical benefits may come from further work based on this design-based approach to consciousness. A deeper longer term implication is the possibility of a new science investigating laws governing possible trajectories in design space and niche space, as these form parts of high order feedback loops in the biosphere.

In this paper the issue of drawing inferences about biological cognitive systems on the basis of connectionist simulations is addressed. In particular, the justification of inferences based on connectionist models trained using the backpropagation learning algorithm is examined. First it is noted that a justification commonly found in the philosophical literature is inapplicable. Then some general issues are raised about the relationships between models and biological systems. A way of conceiving the role of hidden units in connectionist networks is then introduced. This, in combination with an assumption about the way evolution goes about solving problems, is then used to suggest a means of justifying inferences about biological systems based on connectionist research.

Categories and word meanings tend to be interpreted as single entities that are activated in the mind when an individual categorizes the world or uses language. We argue that this view is mistaken by using models based on neural networks and adaptive behavior simulations. Connectionist simulations of populations of organisms that live and reproduce in an environment show that categories and word meanings are open and changing collections of activation patterns in the network's internal units rather than single activation patterns.

A central mystery of visual perception is the classical problem of invariant object recognition: Different appearances of an object can be perceived as ``the same'', despite, e.g., changes in position or illumination, distortions, or partial occlusion by other objects. This article reports on a recent email discussion over the question whether a neural network can learn the simplest of these invariances, i.e. generalize over the position of a pattern on the input layer, including the author's view on what ``learning shift-invariance'' could mean. That definition leaves the problem unsolved. A similar problem is the one of learning to detect symmetries present in an input pattern. It has been solved by a standard neural network requiring some 70000 input examples. Both leave some doubt if backpropagation learning is a realistic model for perceptual processes. Abandoning the view that a stimulus-response system showing the desired behavior must be learned from scratch, yields a radically different solution. Perception can be seen as an active process that rapidly converges from some initial state to an ordered state, which in itself codes for a percept. As an example, I will present a solution to the visual correspondence problem, which greatly alleviates both problems mentioned above.

Analogical inference depends on systematic substitution of the components of compositional structures. Simple systematic substitution has been achieved in a number of connectionist systems that support binding (the ability to create connectionist representations of the combination of component representations). These systems have used various implementations of two generic composition operators: bind() and bundle(). This paper introduces a novel implementation of the bind() operator that is simple, can be efficiently implemented, and highlights the relationship between retrieval queries and analogical mapping. A frame of role/filler bindings can easily be represented using bind() and bundle(). However, typical binding systems are unable to adequately represent multiple frames and arbitrary nested compositional structures. A novel family of representational operators (called braid()) is introduced to address these problems. Other binding systems make the strong assumption that the roles and fillers are disjoint in order to avoid ambiguities inherent in their representational idioms. The braid() operator can be used to avoid this assumption. The new representational idiom suggests how the cognitive processes of bottom-up and top-down object recognition might be implemented. These processes depend on analogical mapping to integrate disjoint representations and drive perceptual search.

A new way of measuring generalization in unsupervised learning is presented. The measure is based on an exclusive allocation, or credit assignment, criterion. In a classifier that satisfies the criterion, input patterns are parsed so that the credit for each input feature is assigned exclusively to one of multiple, possibly overlapping, output categories. Such a classifier achieves context-sensitive, global representations of pattern data. Two additional constraints, sequence masking and uncertainty multiplexing, are described; these can be used to refine the measure of generalization. The generalization performance of EXIN networks, winner-take-all competitive learning networks, linear decorrelator networks, and Nigrin's SONNET-2 network is compared.

The evolution of language implies the parallel evolution of an ability to respond appropriately to signals (language understanding) and an ability to produce the appropriate signals in the appropriate circumstances (language production). When linguistic signals are produced to inform other individuals, individuals that respond appropriately to these signals may increase their reproductive chances but it is less clear what is the reproductive advantage for the languages producers. We present simulations in which populations of neural networks living in an environment evolve a simple language with an informative function. Signals are produced to help other individuals to categorize edible and poisonous mushrooms in order to decide whether to approach or avoid encountered mushrooms. Language production, while not under direct evolutionary pressure, evolves as a by-product of the independently evolving perceptual ability to categorize mushrooms.

Contains extended abstracts of the NIPS*97 workshop "Can Artificial Models Compete to Control Robots?"

Studies of the categorical perception (CP) of sensory continua have a long and rich history in psychophysics. A major development was Macmillan et al.'s application in 1977 of signal detection theory to analyze several experimental paradigms, in particular explicating the relation between the psychometric labeling function and discrimination measures. Simultaneously, Anderson et al. proposed a neural model for what we will term synthetic CP, yet this line of research has been less well explored. In this paper, we assess neural-network models of CP with particular reference to their ability to predict the psychophysical behavior of real observers -- including the relation between labeling and discrimination. Synthetic categorization of a variety of stimuli, including speech sounds and artificial/novel dimensions, is reviewed and discussed in terms of both classical theories of CP and more recent developments. A variety of neural mechanisms is capable of replicating the essentials of categorical perception, indicating that CP is not a special mode of perception but an emergent property of any sufficiently-powerful general learning system. However, the most convincing replication is from a simulation whose output is continuous rather than discrete.

The symbol-based, correspondence epistemology used in AI is contrasted with the constructivist, coherence epistemology promoted by cybernetics. The latter leads to bootstrapping knowledge representations, in which different parts of the cognitive system mutually support each other. Gordon Pask's entailment meshes and their implementation in the ThoughtSticker program are reviewed as a basic application of this methodology. Entailment meshes are then extended to entailment nets: directed graph representations governed by the "bootstrapping axiom", determining which concepts are to be distinguished or merged. This allows a constant restructuring and elicitation of the conceptual network. Semantic networks and frame-like representations with inheritance can be expressed in this very general scheme by introducing a basic ontology of node and link types. Entailment nets are then generalized to associative nets characterized by weighted links. Learning algorithms are presented which can adapt the link strengths, based on the frequency with which links are selected by hypertext browsers. It is argued that these different bootstrapping methods could be applied to make the World-Wide Web more intelligent, by allowing it to self-organize and support inferences through spreading activation.

Intracortical microstimulation (ICMS) of a localized site in the somatosensory cortex of rats and monkeys for 2-6 hours produces a large increase in the cortical representation of the skin region represented by the ICMS-site neurons before ICMS, with very little effect on the ICMS-site neuron's RF location, RF size, and responsiveness (Recanzone et al., 1992). The "EXIN" (afferent excitatory and lateral inhibitory) learning rules (Marshall, 1995) are used to model RF changes during ICMS. The EXIN model produces reorganization of RF topography similar to that observed experimentally. The possible role of inhibitory learning in producing the effects of ICMS is studied by simulating the EXIN model with only lateral inhibitory learning. The model also produces an increase in the cortical representation of the skin region represented by the ICMS-site RF. ICMS is compared to artificial scotoma conditioning (Pettet & Gilbert, 1992) and retinal lesions (Darian-Smith & Gilbert, 1995), and it is suggested that lateral inhibitory learning may be a general principle of cortical plasticity.