Search

Metabolic-repair models, or (M,R)-systems were introduced in Relational Biology by Robert Rosen. Subsequently, Rosen represented such (M,R)-systems (or simply MRs) in terms of categories of sets, deliberately selected without any structure other than the discrete topology of sets. Theoreticians of life’s origins postulated that Life on Earth has begun with the simplest possible organism, called the primordial. Mathematicians interested in biology attempted to answer this important question of the minimal living organism by defining the functional relations that would have made life possible in such a minimal system- a grandad and grandma of all living organisms on Earth.

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/transfer 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.

As a human product, language reflects the psychological experience of man (Radden and Dirven, 2007). One model of language and human cognition in general is connectionism, by many linguists is regarded as mathematical and, therefore, too reductive. This opinion trend seems to be reversing, however, due to the fact that many cognitive researchers begin to appreciate one attribute of network models: feature emergence. In the course of a network simulation properties emerge that were neither inbuilt nor intended by its creators (Elman, 1998), in other words, the whole becomes more than just the sum of its parts. Insight is not only drawn from the network's output, but also the means that the network utilizes to arrive at the output. It may seem obvious that the events of life should be meaningful for human beings, yet there is no widely accepted theory as to how do we derive that meaning. The most promising hypothesis regarding the question how the world is meaningful to us is that of embodied cognition (cf. Turner 2009), which postulates that the functions of the brain evolved so as to ‘understand’ the body, thus grounding the mind in an experiential foundation. Yet, the relationship between the body and the mind is far from perspicuous, as research insight is still intertwined with metaphors specific for the researcher’s methodology (Eliasmith 2003). It is the aim of this paper to investigate the conceptual metaphor in a manner that will provide some insight with regard to the role that objectification, as defined by Szwedek (2002), plays in human cognition and identify one possible consequence of embodied cognition. If the mechanism for concept formation, or categorization of the world, resembles a network, it is reasonable to assume that evidence for this is to be sought in language. Let us then postulate the existence of a network mechanism for categorization and concept formation present in the human mind and initially developed to cope with the world directly accessible to the early human (i.e. tangible). Such a network would convert external inputs to form an internal, multi modal representation of a perceived object in the brain. The sheer amount of available information and the computational restrictions of the brain would force some sort of data compression, or a computational funnel. It has been shown that a visual perception network of this kind can learn to accurately label patterns (Elman, 1998). What is more, the compression of data facilitated the recognition of prototypes of a given pattern category rather than its peripheral representations, an emergent property that supports the prototype theory of the mental lexicon (cf. Radden and Dirven, 2007). The present project proposes that, in the domain of cognition, the process of objectification, as defined by Szwedek (2002), would be an emergent property of such a system, or that if an abstract notion is computed by a neural network designed to cope with tangible concepts the data compression mechanism would require the notion to be conceptualized as an object to permit further processing. The notion of emergence of meaning from the operation of complex systems is recognised as an important process in a number of studies on metaphor comprehension. Feature emergence is said to occur when a non-salient feature of the target and the vehicle becomes highly salient in the metaphor (Utsumi 2005). Therefore, for example, should objectification emerge as a feature in the metaphor KNOWLEDGE IS A TREASURE, the metaphor would be characterised as having more features of an object than either the target or vehicle alone. This paper focuses on providing a theoretical connectionist network based on the Elman-type network (Elman, 1998) as a model of concept formation where objectification would be an emergent feature. This is followed by a psychological experiment whereby the validity of this assumption is tested through a questionnaire where two groups of participants are asked to evaluate either metaphors or their components. The model proposes an underlying relation between the mechanism for concept formation and the omnipresence of conceptual metaphors, which are interpreted as resulting from the properties of the proposed network system. Thus, an evolutionary neural mechanism is proposed for categorization of the world, that is able to cope with both concrete and abstract notions and the by-product of which are the abstract language-related phenomena, i.e. metaphors. The model presented in this paper aims at providing a unified account of how the various types of phenomena, objects, feelings etc. are categorized in the human mind, drawing on evidence from language. References: Szwedek, Aleksander. 2002. Objectification: From Object Perception To Metaphor Creation. In B. Lewandowska-Tomaszczyk and K. Turewicz (eds). Cognitive Linguistics To-day, 159-175. Frankfurt am Main: Peter Lang. Radden, Günter and Dirven, René. 2007. Cognitive English Grammar. Amsterdam/ Philadelphia: John Benjamins Publishing Company Eliasmith, Chris. 2003. Moving beyond metaphors: understanding the mind for what it is. Journal of Philosophy. C(10):493- 520. Elman, J. L. et al. 1998. Rethinking innateness: A connectionist perspective on development. Cambridge, MA: MIT Press Turner, Mark. 2009. Categorization of Time and Space Through Language. (Paper presented at the FOCUS2009 conference "Categorization of the world through language". Serock, 25-28 February 2009). Utsumi, Akira. 2005. The role of feature emergence in metaphor appreciation, Metaphor and Symbol, 20(3), 151-172.

The Line flow Contingency Selection and Ranking (CS & R) is performed to rank the critical contingencies in order of their severity. An Artificial Neural Network based method for MW security assessment corresponding to line outage events have been reported by various authors in the literature. One way to provide an understanding of the behaviour of Neural Networks is to extract rules that can be provided to the user. The domain knowledge (fuzzy rules extracted from Multi-layer Perceptron model trained by Back Propagation algorithm) is integrated into a Neural Network for fast and accurate CS & R in an IEEE 14-bus system, for unknown load patterns and are found to be suitable for on-line applications at Energy Management Centers. The system user is provided with the capability to determine the set of conditions under which a line-outage is critical, and if critical, then how severe it is, thereby providing some degree of transparency of the ANN solution.

The problem of the automatic development of controllers for vehicles for which the exact characteristics are not known is considered in the context of miniature helicopter flocking. A methodology is proposed in which neural network based controllers are evolved in a simulation using a dynamic model qualitatively similar to the physical helicopter. Several network architectures and evolutionary sequences are investigated, and two approaches are found that can evolve very competitive controllers. The division of the neural network into modules and of the task into incremental steps seems to be a precondition for success, and we analyse why this might be so.

This paper addresses the problem of automatically constructing tracks tailor-made to maximize the enjoyment of individual players in a simple car racing game. To this end, some approaches to player modeling are investigated, and a method of using evolutionary algorithms to construct racing tracks is presented. A simple player-dependent metric of entertainment is proposed and used as the fitness function when evolving tracks. We conclude that accurate player modeling poses some significant challenges, but track evolution works well given the right track representation.

Evolutionary car racing (ECR) is extended to the case of two cars racing on the same track. A sensor representation is devised, and various methods of evolving car controllers for competitive racing are explored. ECR can be combined with co-evolution in a wide variety of ways, and one aspect which is explored here is the relative-absolute fitness continuum. Systematical behavioural differences are found along this continuum; further, a tendency to specialization and the reactive nature of the controller architecture are found to limit evolutionary progress.

Neural network-based controllers are evolved for racing simulated R/C cars around several tracks of varying difficulty. The transferability of driving skills acquired when evolving for a single track is evaluated, and different ways of evolving controllers able to perform well on many different tracks are investigated. It is further shown that such generally proficient controllers can reliably be developed into specialized controllers for individual tracks. Evolution of sensor parameters together with network weights is shown to lead to higher final fitness, but only if turned on after a general controller is developed, otherwise it hinders evolution. It is argued that simulated car racing is a scalable and relevant testbed for evolutionary robotics research, and that the results of this research can be useful for commercial computer games.

This paper describes the evolution of controllers for racing a simulated radio-controlled car around a track, modelled on a real physical track. Five different controller architectures were compared, based on neural networks, force fields and action sequences. The controllers use either egocentric (first person), Newtonian (third person) or no information about the state of the car (open-loop controller). The only controller that is able to evolve good racing behaviour is based on a neural network acting on egocentric inputs.

We present here an example of incremental learning between two computational models dealing with different modalities: a model allowing to switch spatial visual attention and a model allowing to learn the ordinal sequence of phonetical numbers. Their merging via a common reward signal allows anyway to produce a cardinal counting behaviour that can be implemented on a robot.

This paper introduces a control architecture for the learning of complex sequence of gestures applied to autonomous robots. The architecture is designed to exploit the robot internal sensory-motor dynamics generated by visual, proprioceptive, and predictive informations in order to provide intuitive behaviors in the purpose of natural interactions with humans.

In this study we investigated the effects of different morphological configurations on a robot swinging task using van der Pol oscillators. The task was examined using two separate degrees of freedom (DoF), both in the presence and absence of neural entrainment. Neural entrainment stabilises the system, reduces time-to-steady state and relaxes the requirement for a strong coupling with the environment in order to achieve mechanical entrainment. It was found that staged release of the distal DoF does not have any benefits over using both DoF from the onset of the experimentation. On the contrary, it is less efficient, both with respect to the time needed to reach a stable oscillatory regime and the maximum amplitude it can achieve. The same neural architecture is successful in achieving neuromechanical entrainment for a robotic walking task.

Image superresolution involves the processing of an image sequence to generate a still image with higher resolution. Classical approaches, such as bayesian MAP methods, require iterative minimization procedures, with high computational costs. Recently, the authors proposed a method to tackle this problem, based on the use of a hybrid MLP-PNN architecture. In this paper, we present a novel superresolution method, based on an evolution of this concept, to incorporate the use of local image models. A neural processing stage receives as input the value of model coefficients on local windows. The data dimension-ality is firstly reduced by application of PCA. An MLP, trained on synthetic se-quences with various amounts of noise, estimates the high-resolution image data. The effect of varying the dimension of the network input space is exam-ined, showing a complex, structured behavior. Quantitative results are presented showing the accuracy and robustness of the proposed method.

Several attempts have been made in the past to construct encoding schemes that allow modularity to emerge in evolving systems, but success is limited. We believe that in order to create successful and scalable encodings for emerging modularity, we first need to explore the benefits of different types of modularity by hard-wiring these into evolvable systems. In this paper we explore different ways of exploiting sensory symmetry inherent in the agent in the simple game Cellz by evolving symmetrically identical modules. It is concluded that significant increases in both speed of evolution and final fitness can be achieved relative to monolithic controllers. Furthermore, we show that a simple function approximation task that exhibits sensory symmetry can be used as a quick approximate measure of the utility of an encoding scheme for the more complex game-playing task.

In this paper we discuss the application of the Kohonen Selforganizing Maps to the classification of triadic chords in inversions and root positions. Our motivation started in the validation of Schönberg´s hypotheses of the harmonic features of each chord inversion. We employed the Kohonen network, which has been generally known as an optimum pattern classification tool in several areas, including music, to verify that hypothesis. The outcomes of our experiment refuse the Schönberg´s assumption in two aspects: structural and perceptual/functional.

The Line flow Contingency Selection and Ranking (CS & R) is performed to rank the critical contingencies in order of their severity. An Artificial Neural Network based method for MW security assessment corresponding to line outage events have been reported by various authors in the literature. One way to provide an understanding of the behaviour of Neural Networks is to extract rules that can be provided to the user. The domain knowledge (fuzzy rules extracted from Multi-layer Perceptron model trained by Back Propagation algorithm) is integrated into a Neural Network for fast and accurate CS & R in an IEEE 14-bus system, for unknown load patterns and are found to be suitable for on-line applications at Energy Management Centers. The system user is provided with the capability to determine the set of conditions under which a line-outage is critical, and if critical, then how severe it is, thereby providing some degree of transparency of the ANN solution.

Binding is one of the most fundamental cognitive functions, how to find the correspondence of sensations between different modalities such as vision and touch. Without a priori knowledge on this correspondence, binding is regarded to be a formidable issue for a robot since it often perceives multiple physical phenomena in its different modal sensors, therefore it should correctly match the foci of attention in different modalities that may have multiple correspondences each other. We suppose that learning the multimodal representation of the body should be the first step toward binding since the morphological constraints in self-body-observation would make the binding problem tractable. The multimodal sensations are expected to be constrained in perceiving own body so as to configurate the unique parts of the multiple correspondence reflecting its morphology. In this paper, we propose a method to match the foci of attention in vision and touch through the unique association by cross-anchoring different modalities. Simple experiments show the validity of the proposed method.

In this paper we propose an approach to robot learning by imitation that uses the multimodal inputs of language, vision and motor. In our approach a student robot learns from a teacher robot how to perform three separate behaviours based on these inputs. We considered two neural architectures for performing this robot learning. First, a one-step hierarchial architecture trained with two different learning approaches either based on Kohonen's self-organising map or based on the Helmholtz machine turns out to be inefficient or not capable of performing differentiated behavior. In response we produced a hierarchial architecture that combines both learning approaches to overcome these problems. In doing so the proposed robot system models specific aspects of learning using concepts of the mirror neuron system (Rizzolatti and Arbib, 1998) with regards to demonstration learning.

We explore two controversial hypotheses through robotic implementation: (1) Processes involved in recognition and response are tightly coupled both in their operation and epigenesis; and (2) processes involved in symbol emergence should respect the integrity of recognition and response while exploiting the periodicity of biological motion. To that end, this paper proposes a method of recognizing and generating motion patterns based on nonlinear principal component neural networks that are constrained to model both periodic and transitional movements. The method is evaluated by an examination of its ability to segment and generalize different kinds of soccer playing activity during a RoboCup match.

We present a quantitative investigation on the effects of a discrete developmental progression on the acquisition of a foveation behavior by a robotic hand-arm-eyes system. Development is simulated by (a) increasing the resolution of visual and tactile systems, (b) freezing and freeing mechanical degrees of freedom, and (c) adding neuronal units to the neural control architecture. Our experimental results show that a system starting with a low-resolution sensory system, a low precision motor system, and a low complexity neural structure, learns faster that a system which is more complex at the beginning.

Denoising source separation is a recently introduced framework for building source separation algorithms around denoising procedures. Two developments are reported here. First, a new scheme for accelerating and stabilising convergence by controlling step sizes is introduced. Second, a novel signal-variance based denoising function is proposed. Estimates of variances of different source are whitened which actively promotes separation of sources. Experiments with artificial data and real magnetoencephalograms demonstrate that the developed algorithms are accurate, fast and stable.

A neural network model is proposed that forms sparse spatiotemporal memory traces of spatiotemporal events given single occurrences of the events. The traces are distributed in that each individual cell and synapse participates in numerous traces. This sharing of representational substrate provides the basis for similarity-based generalization and thus semantic memory. Simulation results are provided demonstrating that similar spatiotemporal patterns map to similar traces. The model achieves this property by measuring the degree of match, G, between the current input pattern on each time slice and the expected input given the preceding time slices (i.e., temporal context) and then adding an amount of noise, inversely proportional to G, to the process of choosing the internal representation for the current time slice. Thus, if G is small, indicating novelty, we add much noise and the resulting internal representation of the current input pattern has low overlap with any preexisting representations of time slices. If G is large, indicating a familiar event, we add very little noise resulting in reactivation of all or most of the preexisting representation of the input pattern.

MISEP has been proposed as a generalization of the INFOMAX method in two directions: (1) handling of nonlinear mixtures, and (2) learning the nonlinearities to be used at the outputs, making the method suitable to the separation of components with a wide range of statistical distributions. In all implementations up to now, MISEP had used multilayer perceptrons (MLPs) to perform the nonlinear ICA operation. Use of MLPs sometimes leads to a relatively slow training. This has been attributed, at least in part, to the non-local character of the MLP's units. This paper investigates the possibility of using a network of radial basis function (RBF) units for performing the nonlinear ICA operation. It shows that the local character of the RBF network's units allows a significant speedup in the training of the system. The paper gives a brief introduction to the basics of the MISEP method, and presents experimental results showing the speed advantage of using an RBF-based network to perform the ICA operation.

By constructing and analyzing a physically situated brain-based device (i.e. a device with sensors and actuators whose behavior is guided by a simulated nervous system), we show that reentrant connectivity and dynamic synchronization can provide an effective mechanism for binding the visual features of objects.

We present the architecture and distributed algorithms of an implemented system called NeuSter, that unifies learning, perception and action for autonomous robot control. NeuSter comprises several sub-systems that provide online learning for networks of million neurons on machine clusters. It extracts information from sensors, builds its own representations of the environment in order to learn non-predefined goals.

We present in this paper a biologically inspired model of the Basal Ganglia which deals with Block Copying as a sequence learning task. By breaking a relatively complex task into simpler operations with well-defined skills, an approach which is termed as a skill-based machine design is used in the device of computational models to complete such tasks. Basal Ganglia are critically involved in sensorimotor control. From the learning aspects, Actor-Critic architectures have been proposed to model the Basal Ganglia and Temporal difference has been proposed as a learning algorithm. The model is implemented and simulation results are presented to show the capability of our model to successfully complete the task.

A connectionist model consisting of thirty cortico-basal ganglionic loops was implemented. This model encodes temporal information into a spatial pattern of neuronal activations in the prefrontal cortex using neurophysiologically plausible activation functions and circuitry without learning. This neural architecture was used to model experiments with infants. Initial results suggest that the cortical basal ganglionic circuitry has an inherent ability to differentiate sequential information.

Historically, a lot of authors in psychology and in robotics tend to separate "true imitation" and its related high-level mechanisms which seem to be exclusive to human adult, from low-level imitations or "mimicries" observed on babies or primates. Closely, classical researches suppose that an imitative artificial system must be able to build a model of the demonstrator's geometry, in order to reproduce finely the movements on each joints. Conversely, we will advocate that if imitation is viewed as a part of a developmental course, then (1) an artificial developing system does not need to build any internal model of the other, to perform real-time and low-level imitations of human movements despite the related correspondence problem between man and robot and, (2) a simple sensory-motor loop could be at the basis of multiples heterogeneous imitative behaviors often explained in the literature by different models.

Given the importance of using resources efficiently in the competition for survival, it is reasonable to think that natural evolution has discovered efficient cortical coding strategies for representing natural visual information. Sparse representations have intrinsic advantages in terms of fault-tolerance and low-power consumption potential, and can therefore be attractive for robot sensorimotor control with powerful dispositions for decision-making. Inspired by the mammalian brain and its visual ventral pathway, we present in this paper a hierarchical sparse coding network architecture that extracts visual features for use in sensorimotor control. Testing with natural images demonstrates that this sparse coding facilitates processing and learning in subsequent layers. Previous studies have shown how the responses of complex cells could be sparsely represented by a higher-order neural layer. Here we extend sparse coding in each network layer, showing that detailed modeling of earlier stages in the visual pathway enhances the characteristics of the receptive fields developed in subsequent stages. The yield network is more dynamic with richer and more biologically plausible input and output representation.

In this paper, we investigated interactive learning between human subjects and robot experimentally, and its essential characteristics are examined using the dynamical systems approach. Our research concentrated on the navigation system of a specially developed humanoid robot called Robovie and seven human subjects whose eyes were covered, making them dependent on the robot for directions. We compared the usual feed-forward neural network (FFNN) without recursive connections and the recurrent neural network (RNN). Although the performances obtained with both the RNN and the FFNN improved in the early stages of learning, as the subject changed the operation by learning on its own, all performances gradually became unstable and failed. Results of a questionnaire given to the subjects confirmed that the FFNN gives better mental impressions, especially from the aspect of operability. When the robot used a consolidation-learning algorithm using the rehearsal outputs of the RNN, the performance improved even when interactive learning continued for a long time. The questionnaire results then also confirmed that the subject's mental impressions of the RNN improved significantly. The dynamical systems analysis of RNNs support these differences and also showed that the collaboration scheme was developed dynamically along with succeeding phase transitions.

The Double Cone Model (DCM) is a model of how the brain transforms sensory input to motor commands through successive stages of data compression and expansion. We have tested a subset of the DCM on speech recognition, production and imitation. The experiments show that the DCM is a good candidate for an artificial speech processing system that can develop autonomously. We show that the DCM can learn a repertoire of speech sounds by listening to speech input. It is also able to link the individual elements of speech to sequences that can be recognized or reproduced, thus allowing the system to imitate spoken language.

The corticothalamic feedback and the thalamic reticular nucleus have gained much attention lately because of their integrative and modulatory functions. A previous study by the author suggested that this circuitry can process analogies (i.e., the {\em analogy hypothesis}). In this paper, the proposed model was implemented as a network of leaky integrate-and-fire neurons to test the {\em analogy hypothesis}. The previous proposal required specific delay and temporal dynamics, and the implemented network tuned accordingly functioned as predicted. Furthermore, these specific conditions turn out to be consistent with experimental data, suggesting that a further investigation of the thalamocortical circuit within the {\em analogical framework} may be worthwhile.

The corticothalamic feedback and the thalamic reticular nucleus have gained much attention lately because of their integrative and modulatory functions. A previous study by the author suggested that this circuitry can process analogies (i.e., the {\em analogy hypothesis}). In this paper, the proposed model was implemented as a network of leaky integrate-and-fire neurons to test the {\em analogy hypothesis}. The previous proposal required specific delay and temporal dynamics, and the implemented network tuned accordingly functioned as predicted. Furthermore, these specific conditions turn out to be consistent with experimental data, suggesting that a further investigation of the thalamocortical circuit within the {\em analogical framework} may be worthwhile.

MISEP has been proposed as a generalization of the INFOMAX method in two directions: (1) handling of nonlinear mixtures, and (2) learning the nonlinearities to be used at the outputs, making the method suitable to the separation of components with a wide range of statistical distributions. In all implementations up to now, MISEP had used multilayer perceptrons (MLPs) to perform the nonlinear ICA operation. Use of MLPs sometimes leads to a relatively slow training. This has been attributed, at least in part, to the non-local character of the MLP's units. This paper investigates the possibility of using a network of radial basis function (RBF) units for performing the nonlinear ICA operation. It shows that the local character of the RBF network's units allows a significant speedup in the training of the system. The paper gives a brief introduction to the basics of the MISEP method, and presents experimental results showing the speed advantage of using an RBF-based network to perform the ICA operation.

Human sound systems are invariably phonemically coded. Furthermore, phoneme inventories follow very particular tendancies. To explain these phenomena, there existed so far three kinds of approaches : ``Chomskyan''/cognitive innatism, morpho-perceptual innatism and the more recent approach of ``language as a complex cultural system which adapts under the pressure of efficient communication''. The two first approaches are clearly not satisfying, while the third, even if much more convincing, makes a lot of speculative assumptions and did not really bring answers to the question of phonemic coding. We propose here a new hypothesis based on a low-level model of sensory-motor interactions. We show that certain very simple and non language-specific neural devices allow a population of agents to build signalling systems without any functional pressure. Moreover, these systems are phonemically coded. Using a realistic vowel articulatory synthesizer, we show that the inventories of vowels have striking similarities with human vowel systems.

This paper proposes an explanation of the cognitive change that occurs as the creative process proceeds. During the initial, intuitive phase, each thought activates, and potentially retrieves information from, a large region containing many memory locations. Because of the distributed, content-addressable structure of memory, the diverse contents of these many locations merge to generate the next thought. Novel associations often result. As one focuses on an idea, the region searched and retrieved from narrows, such that the next thought is the product of fewer memory locations. This enables a shift from association-based to causation-based thinking, which facilitates the fine-tuning and manifestation of the creative work.

Given a neural control structure, what would be the impact of body growth on control performance? This question, which addresses the issue of the interaction between innate structure, ongoing developing structure and experience, is very relevant to the field of epigenetic robotics. Much of the early social interaction is done as the body develops and the interplay cannot be ignored. We hypothesize that starting with fewer degrees of freedom enables a more efficient exploration of the sensorimotor space, that results in multiple directions of stability. While not necessarily corresponding to optimal task performance, they will guide the coordination of additional degrees of freedom. These additional degrees of freedom then allow for optimal task performance as well as for more tolerance and adaptation to environmental interaction. We propose a simple case-study to validate our hypothesis and describe experiments with a small humanoid robot.

We present the first stages of the developmental course of a robot using vision and a 5 degree of freedom robotic arm. During an exploratory behavior, the robot learns visuo-motor control of its mechanical arm. We show how a simple neural network architecture, combining elementary vision, a self-organized algorithm, and dynamical Neural Fields is able to learn and use proper associations between vision and arm movements, even if the problem is ill posed (2-D toward 3-D mapping and also mechanical redundancy between different joints). Highlighting the generic aspect of such an architecture, we show as a robotic result that it is used as a basis for simple gestural imitations of humans. Finally we show how the imitative mechanism carries on the developmental course, allowing the acquisition of more and more complex behavioral capabilities.

A functional focal-plane implementation of a 2D optical flow system is presented that detects an preserves motion discontinuities. The system is composed of two different network layers of analog computational units arranged in a retinotopical order. The units in the first layer (the optical flow network) estimate the local optical flow field in two visual dimensions, where the strength of their nearest-neighbor connections determines the amount of motion integration. Whereas in an earlier implementation \cite{Stocker_Douglas99} the connection strength was set constant in the complete image space, it is now \emph{dynamically and locally} controlled by the second network layer (the motion discontinuities network) that is recurrently connected to the optical flow network. The connection strengths in the optical flow network are modulated such that visual motion integration is ideally only facilitated within image areas that are likely to represent common motion sources. Results of an experimental aVLSI chip illustrate the potential of the approach and its functionality under real-world conditions.

Shepard and Chipman's second order isomorphism describes how the brain may represent the relations in the world. However, a common interpretation of the theory can cause difficulties. The problem originates from the static nature of representations. In an alternative interpretation, I propose that we assign an active role to the internal representations and relations. It turns out that a collection of such active units can perform analogical tasks. The new interpretation is supported by the existence of neural circuits that may be implementing such a function. Within this framework, perception, cognition, and motor function can be understood under a unifying principle of analogy.

This paper presents an Artificial Life and Neural Network (ALNN) model for the evolution of syntax. The simulation methodology provides a unifying approach for the study of the evolution of language and its interaction with other behavioral and neural factors. The model uses an object manipulation task to simulate the evolution of language based on a simple verb-noun rule. The analyses of results focus on the interaction between language and other non-linguistic abilities, and on the neural control of linguistic abilities. The model shows that the beneficial effects of language on non-linguistic behavior are explained by the emergence of distinct internal representation patterns for the processing of verbs and nouns.

This paper describes a neural network model for the study of spatial language. It deals with both geometric and functional variables, which have been shown to play an important role in the comprehension of spatial prepositions. The network is integrated with a virtual reality interface for the direct manipulation of geometric and functional factors. The training uses experimental stimuli and data. Results show that the networks reach low training and generalization errors. Cluster analyses of hidden activation show that stimuli primarily group according to extra-geometrical variables.

In this paper we present a self-organizing connectionist model of the acquisition of word meaning. Our model consists of two neural networks and builds on the basic concepts of Hebbian learning and self-organization. One network learns to approximate word transition probabilities, which are used for lexical representation, and the other network, a self-organizing map, is trained on these representations, projecting them onto a 2D space. The model relies on lexical co-occurrence information to represent word meanings in the lexicon. The results show that our model is able to acquire semantic representations from both artificial data and real corpus of language use. In addition, the model demonstrates the ability to develop rather accurate word representations even with a sparse training set.

We examined the influence of the pre-existing organiza-tion of the semantic memory on forming new episodic associations between words. Testing human subjects' performance we found that a semantic relationship be-tween words facilitates forming episodic associations be-tween them. Furthermore, the amount of facilitation in-creases linearly as a function of the number of co-occurrence of the words, up to a ceiling. Constrained by these empirical findings we developed a computational model, based on the theory of spreading activation over semantic networks. The model uses self-organizing maps to represent semantic relatedness, and lateral connections to represent the episodic associations. When two words are presented to the model, the interaction of the two ac-tivation waves is summed and added to the direct lateral connection between them. The main result is that the model is capable of replicating the empirical results. The model also makes several testable predictions: First, it should be easier to form an association from a word with few semantic neighbors to a word with many se-mantic neighbors than vice-versa. Second, after associat-ing an unrelated word pair it should be easier to associate another two words each related to one of the words in the first pair. Third, a less focused activation wave, which may be the cause of schizophrenic thought disor-der, should decrease the advantage in learning rate of re-lated over unrelated pairs.

Neural networks that learn the What and Where task perform better if they possess a modular architecture for separately processing the identity and spatial location of objects. In previous simulations the modular architecture either was hardwired or it developed during an individual's life based on a preference for short connections given a set of hardwired unit locations. We present two sets of simulations in which the network architecture is genetically inherited and it evolves in a population of neural networks in two different conditions: (1) both the architecture and the connection weights evolve; (2) the network architecture is inherited and it evolves but the connection weights are learned during life. The best results are obtained in condition (2). Condition (1) gives unsatisfactory results because (a) adapted sets of weights can suddenly become maladaptive if the architecture changes, (b) evolution fails to properly assign computational resources (hidden units) to the two tasks, (c) genetic linkage between sets of weights for different modules can result in a favourable mutation in one set of weights being accompanied by an unfavourable mutation in another set of weights.

It is argued that cognitive capacities can be understood as the outcome of the collective action of a set of agents created by tools that explore possible behaviours and train the agents to behave in such appropriate ways as may be discovered. The coherence of the whole system is assured by a combination of vetting the performance of new agents and dealing appropriately with any faults that the whole system may develop. This picture is shown to account for a range of cognitive capacities, including language.

from Introduction: Animal communication systems and human languages can be characterised by the type of cognitive abilities that are required. If we consider the main semiotic distinction between communication using icons, signals, or symbols (Peirce, 1955; Harnad, 1990; Deacon, 1997) we can identify different cognitive loads for each type of reference. The use and understanding of icons require instinctive behaviour (e.g. emotions) or simple perceptual processes (e.g. visual similarities between an icon and its meaning). Communication systems that use signals are characterised by referential associations between objects and visual or auditory signals. They require the cognitive ability to learn stimulus associations, such as in conditional learning. Symbols have double associations. Initially, symbolic systems require the establishment of associations between signals and objects. Secondly, other types of relationships are learned between the signals themselves. The use of rule for the logical combination of symbols is an example of symbolic relationship. Symbolisation is the ability to acquire and handle symbols and symbolic relationships.