Search

We introduce a general method to extract knowledge from a recurrent neural network (Long Short Term Memory) that has learnt to detect if a given input sequence is valid or not, according to an unknown generative automaton. Based on the clustering of the hidden states, we explain how to build and validate an automaton that corresponds to the underlying (unknown) automaton, and allows to predict if a given sequence is valid or not. The method is illustrated on artificial grammars (Reber's grammar variations) as well as on a real use-case whose underlying grammar is unknown., 18 pages, 17 figures

International audience; Complex problem solving involves representing structured knowledge, reasoning and learning, all at once. In this prospective study, we make explicit how a reinforcement learning paradigm can be applied to a symbolic representation of a concrete problem-solving task, modeled here by an ontology. This preliminary paper is only a set of ideas while feasibility verification is still a perspective of this work.

A critical understanding of digital technologies is an empowering competence for citizens of all ages. In this paper we introduce an open educational approach of artificial intelligence (AI) for everyone. Through a hybrid and participative MOOC we aim to develop a critical and creative perspective about the way AI is integrated in the different domains of our lives. We have built and now operate a MOOC in AI for all the citizens from 15 years old. The MOOC aims to help understanding AI foundations and applications, intended for a large public beyond the school domain, with more than 20,000 participants engaged in the MOOC after nine months. This study addresses the pedagogical methods for designing and evaluating the MOOC in AI. Through this study we raise four questions regarding citizen education in AI: Why (i.e., to which aim) sharing such citizen formation? What is the disciplinary knowledge to be shared? What are the competencies to develop? How can it be shared and evaluated? We finally share learning analytics, quantitative and qualitative evaluations and explain to which extent educational science research helps enlighten such large scale initiatives. The analysis of the MOOC in AI helps to identify that the main feedback related to AI is “fear”, because AI is unknown and mysterious to the participants. After developing playful AI simulations, the AI mechanisms become familiar for the MOOC participants and they can overcome their misconception on AI to develop a more critical point of view. This contribution describes a K-12 AI educational project or initiatives of a considerable impact, via the formation of teachers and other educators. Supplementary Information The online version contains supplementary material available at 10.1007/s13218-021-00725-7.

A major cognitive function is often overlooked in artificial intelligence research: episodic memory. In this paper, we relate episodic memory to the more general need for explicit memory in intelligent processing. We describe its main mechanisms and its involvement in a variety of functions, ranging from concept learning to planning. We set the basis for a computational cognitive neuroscience approach that could result in improved machine learning models. More precisely, we argue that episodic memory mechanisms are crucial for contextual decision making, generalization through consolidation and prospective memory.

International audience; This paper presents a new approach for learning transition function in state representation learning (SRL) for control. While state-of-the-art methods use different deterministic neural networks to learn forward and inverse state transition functions independently with auto-supervised learning, we introduce a bidirectional stochastic model to learn both transition functions. We aim at using the uncertainty of the model on its predictions as an intrinsic motivation for exploration to enhance the representation learning. More, using the same model to learn both transition functions allows sharing the parameters, which can reduce their number and should increase the embedding quality of the representation. We use a factored restricted Boltzmann machine (fRBM) based model, enhanced with dedicated structure for learning system dynamics and transitions with shared parameters. The presented work focuses on building the structure of the bidirectional transition model for unsupervised learning. Our fRBM structure is directly inspired from physics interactions between inputs and outputs in reinforcement learning framework. We compare different training algorithms for learning the model that must be able to predict observable random variables to be used in SRL framework. Our structure is not restricted to any type of observable, nevertheless in this paper we focus on learning dynamics from the OpenAI Gym environment Swinging Pendulum. We show that the proposed structure is able to learn bidirectional transition function and performs well in prediction task.

Computational Intelligence and Artificial Intelligence are both aiming at building machines and softwares capable of intelligent behavior. They are consequently prone to interactions, even if the latter is not necessarily interested in understanding how cognition emerges from the brain substrate. In this chapter, we enumerate, describe and discuss the most important fields of interactions. Some are methodological and are concerned with information representation, processing and learning. At the functional level, the focus is set on major cognitive functions like perception, navigation, decision making and language. Among the salient characteristics of the critical contributions of Computational Neuroscience to the development of intelligent systems, its systemic view of the cerebral functioning is particularly precious to model highly multimodal cognitive functions like decision making and language and to design cognitive architectures for the autonomous behavior of robots.

International audience; Pattern separation is a fundamental hippocampal process thought to be critical for distinguishing similar episodic memories, and has long been recognized as a natural function of the dentate gyrus (DG) supporting autoassociative learning in CA3. Understanding how neural circuits within the DG-CA3 network mediate this process has received much interest, yet the exact mechanisms behind remain elusive. Here we argue for the case that sparse coding is necessary but not sufficient to ensure efficient separation and, alternatively, propose a possible interaction of distinct circuits which, nevertheless, act in synergy to produce a unitary function of pattern separation. The proposed circuits involve different functional granule-cell populations, a primary population mediates sparsification and provides recurrent excitation to the other populations which are related to additional pattern separation mechanisms with higher degrees of robustness against interference in CA3. A variety of top-down and bottom-up factors, such as motivation, emotion, and pattern similarity, controls the selection of circuitry depending on circumstances. According to this framework, a computational model is implemented and tested against model variants in a series of numerical simulations and biological experiments. The results demonstrate that the model combines fast learning, robust pattern separation and high storage capacity. It also accounts for the controversy around the involvement of the DG during memory recall, explains other puzzling findings, and makes predictions that can inform future investigations.

In a context of agricultural intensification and increasing urbanization, the biodiversity of farmed landscapes is a key to improve the sustainability of agro-ecosystems. We seek to ascertain the plant biodiversity of farmed and abandoned vineyards and olive-groves and to identify the factors underlying it: natural and cultural; on local, landscape, and regional scales. To do so, we recorded and calculated the floral biodiversity of 106 georeferenced plots and 121 plot edges distributed across 6 French Mediterranean terroirs, surveyed the practices and perception of 55 farmers, and mapped the landscapes of 20 communes in a GIS. Statistical tests proved that richness and spatial diversity on plots are favored by local low intensity management integrating heritage and landscape objectives. The presence of edges augments the richness and diversity around vineyards. The highest value of spatial diversity was found using the terroir variable. Maximum richness is found in olive groves which are maintained by amateur gardeners and located in the middle of the urbanization gradient. The diversity of biological traits is listed according to: (a) an herbaceous diversity gradient explained by management; (b) a specialization gradient explained by landscapes and distance to large urban areas. Our results draw perspectives to improve existing models of the links between agriculture, biodiversity, and landscape, considering cultural and geographical factors. They lead to recommendations regarding the management of landscapes based on local knowledge and good practices.

International audience; We discuss here the characteristics of a software environment appropriate for the development of a bio-inspired cognitive architecture, which can emulate the behavior of autonomous intelligent agents. First, it is reminded that, while the focus is often set on the more abstract aspects of cognitive abilities, studying the fundamental bases of intelligence that allow for autonomy is a prerequisite for well defined intelligent systems. Secondly, we highlight functional loops associating cerebral structures including the basal ganglia in the brain of most species along the evolution. They are dedicated to the organization of behavior under the constraint of reinforcement , corresponding in their simplest expression to the selection of action for survival. Lastly, concerning the simulation of such models, we describe a software environment to study such relations in a more controlled way than hardware implementations, by adapting a platform built on the top of a video game for the development of classical artificial intelligence models. We explain here how our neuronal model exhibits bodily and internal characteristics necessary for survival tasks and how these characteristics are plugged in the simulation platform. Some scenarios of survival are reported as an illustration of this environment.

International audience; The neurocomputational model described here proposes that two dimensions involved in computation of reward prediction errors i.e magnitude and time could be computed separately and later combined unlike traditional reinforcement learning models. The model is built on biological evidences and is able to reproduce various aspects of classical conditioning, namely, the progressive cancellation of the predicted reward, the predictive firing from conditioned stimuli, and delineation of early rewards by showing firing for sooner early rewards and not for early rewards that occur with a longer latency in accordance with biological data.

The electrical domain, either domestic or industrial, benefits from a huge set of well-defined norms at both the national and international levels. However and surprisingly enough, there is no such norm regarding the actual conception and structuration of electrical diagrams, even though the basic symbols and notations remain the same. Each company is actually free to design such diagram relative to its own experience , expertise and know-how. The difficulty is that such diagrams, which are most of the time materialized as a PDF booklet, do not reflect this implicit knowledge. In this paper, we introduce our work on the extraction and the structuration of such knowledge using ad-hoc graph and text analysis as well as clustering techniques. Starting from a set of raw documents, we propose an end-to-end solution that offers a structured view, which is company dependent, of any electrical diagram for a given company.

In real learning paradigms like pavlovian conditioning, several modes of learning are associated, including generalization from cues and integration of specific cases in their context. Associative memories have been shown to be interesting neuronal models to learn quickly specific cases but they are hardly used in realistic applications because of their limited storage capacities resulting in interference when too many examples are considered. Inspired by biological considerations, we propose a modular model of associative memory including mechanisms to manipulate properly multimodal inputs and to detect and manage interference. This paper reports experiments that demonstrate the good behavior of the model in a wide series of simulations and discusses its impact both in machine learning and in biological modeling.

International audience; Recently, Machine Learning has achieved impressive results, surpassing human performances, but these powerful algorithms are still unable to define their goals by themselves or to adapt when the task changes. In short, they are not autonomous. In this paper, we explain why autonomy is an important criterion for really powerful learning algorithms. We propose a number of characteristics that make humans more autonomous than machines when they learn. Humans have a system of memories where one memory can compensate or train another memory if needed. They are able to detect uncertainties and adapt accordingly. They are able to define their goals by themselves, from internal and external cues and are capable of self-evaluation to adapt their learning behavior. We also suggest that introducing these characteristics in the domain of Machine Learning is a critical challenge for future intelligent systems.

International audience; Noradrenaline participates in the neuromodulation of brain activity to modify the trade-off between exploration and exploitation when sensory contingencies have changed. Accordingly, attentional models of noradrenaline acting on sensory representations have been proposed. In this paper , we explore another possible action of this neuromodulator in the decision making process and report simulation results that illustrate that its role is concerned with different aspects of sensory processing. This is made possible by the extension of a classical model of action selection, to render it able to detect and to adapt to sudden changes in sensory contingencies, which is a major characteristic of autonomous learning.

International audience; Neuromodulation is an interesting way to display different modes of functioning in a complex network. The effect of Noradrenaline has often been related to the exploration/exploitation trade-off and implemented in models by modulation of the gain of activation function. In this paper, we show that this mechanism is not sufficient for system-level networks and propose another way to implement it, exploiting reported inhibition of a striatal region by Noradrenaline. We describe here the corresponding model and report its performances in a reversal task.

International audience; One of the most critical properties of a versatile intelligent agent is its capacity to adapt autonomously to any change in the environment without overly complexifying its cognitive architecture. In this paper, we propose that understanding the role of neuromodulation in the brain is of central interest for this purpose. More precisely, we propose that an accurate estimation of the nature of uncertainty present in the environment is performed by specific brain regions and broadcast throughout the cerebral network by neuromodulators, resulting in appropriate changes in cerebral functioning and learning modes. Better understanding the principles of these mechanisms in the brain might tremendously inspire the field of Artificial General Intelligence. The original contribution of this paper is to relate the four major neuromodulators to four fundamental dimensions of uncertainty.

Most conventional learning algorithms require both positive and negative training data for achieving accurate classification results. However, the problem of learning classifiers from only positive data arises in many applications where negative data are too costly, difficult to obtain, or not available at all. This paper describes a new machine learning approach, called ILoNDF (Incremental data-driven Learning of Novelty Detector Filter). The approach is inspired by novelty detection theory and its learning method, which typically requires only examples from one class to learn a model. One advantage of ILoNDF is the ability of its generative learning to capture the intrinsic characteristics of the training data by continuously integrating the information relating to the relative frequencies of the features of training data and their co-occurrence dependencies. This makes ILoNDF rather stable and less sensitive to noisy features which may be present in the representation of the positive data. In addition, ILoNDF does not require extensive computational resources since it operates on-line without repeated training, and no parameters need to be tuned. In this study we mainly focus on the robustness of ILoNDF in dealing with high-dimensional noisy data and we investigate the variation of its performance depending on the amount of data available for training. To make our study comparable to previous studies, we investigate four common methods: PCA residuals, Hotelling's T 2 test, an auto-associative neural network, and a one-class version of the SVM classifier (lately a favored method for one-class classification). Experiments are conducted on two real-world text corpora: Reuters and WebKB. Results show that ILoNDF tends to be more robust, is less affected by initial settings, and consistently outperforms the other methods.

We ask the question whether the coding of categorical versus coordinate spatial relations depends on different neural networks showing hemispheric specialization or whether there is continuity between these two coding types. The 'continuous spatial coding' hypothesis would mean that the two coding types rely essentially on the same neural network consisting of more general-purpose processes, such as visuo-spatial attention, but with a different weighting of these general processes depending on exact task requirements. With event-related fMRI, we have studied right-handed male subjects performing a grid/no-grid visuo-spatial working memory task inducing categorical and coordinate spatial relations coding. Our data support the 'continuous spatial coding' hypothesis, indicating that, while based on the same fronto-parieto-occipital neural network than categorical spatial relations coding, the coding of coordinate spatial relations relies more heavily on attentional and executive processes, which could induce hemispheric differences similar to those described in the literature. The results also show that visuo-spatial working memory consists of a short-term posterior store with a capacity of up to three elements in the parietal and extrastriate cortices. This store depends on the presence of a visible space categorization and thus can be used for the coding of categorical spatial relations. When no visible space categorization is given or when more than three elements have to be coded, additional attentional and executive processes are recruited, mainly located in the dorso-lateral prefrontal cortex.

International audience; Speech perception and recognition using biologically inspired models is a challenging issue not well explored yet. The paper presents two spatio-temporal biologically inspired methods for the preprocessing and learning of speech signal based on self-organization map (SOM). The experimental results are very encouraging and provide a framework that could be more studied to enhance speech perception and recognition technology.

Best Paper Award; International audience; Neuronal models of associative memories are recurrent networks able to learn quickly patterns as stable states of the network. Their main acknowledged weakness is related to catastrophic interference when too many or too close examples are stored. Based on biological data we have recently proposed a model resistant to some kinds of interferences related to heteroassociative learning. In this paper we report numerical experiments that highlight this robustness and demonstrate very good performances of memorization. We also discuss convergence of interests for such an adaptive mechanism for biological modeling and information processing in the domain of machine learning.

International audience; Studying and modeling the brain as a whole is a real challenge. For such systemic models (in contrast to models of one brain area or aspect), there is a real need for new tools designed to perform complex numerical experiments, beyond usual tools distributed in the computer science and neuroscience communities. Here, we describe an effective solution, freely available on line and already in use, to validate such models of the brain functions. We explain why this is the best choice, as a complement to robotic setup, and what are the general requirements for such a benchmarking platform. In this experimental setup, the brainy-bot implementing the model to study is embedded in a simplified but realistic controlled environment. From visual, tactile and olfactory input, to body, arm and eye motor command, in addition to vital interoceptive cues, complex survival behaviors can be experimented. We also discuss here algorithmic high-level cognitive modules, making the job of building biologically plausible bots easier. The key point is to possibly alternate the use of symbolic representation and of complementary and usual neural coding. As a consequence, algorithmic principles have to be considered at higher abstract level, beyond a given data representation, which is an interesting challenge.

International audience; Artificial Neural Networks are often used as black boxes to implement behavioral functions, developed by trials and errors, fed with sensory inputs and controlled by some criteria of performance. This is the case for pavlovian conditioning where important sensory information is non ambiguous and where the error of prediction is to be minimized. These past years, taking into account critical conditioning behaviors entailed complexifying the neuronal functioning and learning rules. This resulted in networks still simple at the architectural level but with a dynamics difficult to master. Instead, we propose a new neuronal model using uniform and classical neuronal dynamics, with a more complex architecture based on recent findings in neuroscience. Results reported in this paper confirm the good behavior of the model and justify the complex architecture by the greater robustness and flexibility of the model.

This study is concerned with understanding of the formation of ore deposits (precious and base metals) and contributes to the exploration and discovery of new occurrences using artificial neural networks. From the different digital data sets available in BRGM's GIS Andes (a comprehensive metallogenic continental-scale Geographic Information System) 25 attributes are identified as known factors or potential factors controlling the formation of gold deposits in the Andes Cordillera. Various multilayer perceptrons were applied to discriminate possible ore deposits from barren sites. Subsequently, because artificial neural networks can be used to construct a revised model for knowledge extraction, the optimal brain damage algorithm by LeCun was applied to order the 25 attributes by their relevance to the classification. The approach demonstrates how neural networks can be used efficiently in a practical problem of mineral exploration, where general domain knowledge alone is insufficient to satisfactorily model the potential controls on deposit formation using the available information in continent-scale information systems.

Article dans revue scientifique avec comité de lecture.; International audience; Artificial neural networks are efficient for performing signal processing but are not able to explain their decision nor to extract knowledge from data. We propose here a way to extract rules and hints from the hidden layers of a multilayered perceptron. The network is first pruned and then the progressive use of a simpler transfer function can allow such knowledge extraction. This method has been successfully applied to radar signal identification.

International audience; Artificial Neural Networks are very efficient adaptive models but one of their recognized weaknesses is about information representation, often carried out in an input vector without a structure. Beyond the classical elaboration of a hierarchical representation in a series of layers, we report here inspiration from neuroscience and argue for the design of heterogenous neural networks, processing information at feature, configuration and history levels of granularity, and interacting very efficiently for high-level and complex decision making. This framework is built from known characteristics of the sensory cortex, the hippocampus and the prefrontal cortex and is exemplified here in the case of pavlovian conditioning, but we propose that it can be advantageously applied in a wider extent, to design flexible and versatile information processing with neuronal computation.

Contents: Preface. R. Sun, An Introduction to Hybrid Connectionist-Symbolic Models. Part I: Reviews and Overviews M. Hilario, An Overview of Strategies for Neurosymbolic Integration. R. Khosla, T. Dillon, Task Structure Level Symbolic-Connectionist Architecture. Y. Lallement, F. Alexandre, Cognitive Aspects of Neurosymbolic Integration. L. Magdalena, A First Approach to a Taxonomy of Fuzzy-Neural Systems. Part II: Learning in Multi-Module Systems T. Johnson, J. Zhang, H. Wang, A Hybrid Learning Model of Abductive Reasoning. R. Sun, T. Peterson, A Hybrid Agent Architecture for Reactive Sequential Decision Making. M. Malek, B. Amy, A Preprocessing Model for Integrating Case-Based Reasoning and Prototype-Based Neural Network. B. Orsier, A. Labbi, A Step Toward Fully Integrated Hybrid Systems. J.C. Gonzalez, J.R. Velasco, C.A. Iglesias, A Distributed Platform for Symbolic-Connectionist Interoperation. Part III: Representing Symbolic Knowledge. B. Kokinov, Micro-Level Hybridization in the Cognitive Architecture Dual. S. Stevenson, An Integrated Symbolic/Connectionist Parsing Architecture. X. Wu, M. McTear, P. Ojha, H. Dai, A Hybrid System Framework for Disambiguating Word Senses. N.S. Park, D. Robertson, A Localist Network Architecture for Logical Inference. J. Austin, A Distributed Associative Memory for Symbolic Reasoning. E. Mjolsness, Symbolic Neural Networks Derived From Stochastic Grammar Domain Models. Part IV: Acquiring Distributed Representation. T.A. Plate, Structure Matching and Transformation With Distributed Representations. A. Sperduti, A. Starita, C. Goller, Distributed Representations for Terms in Hybrid Reasoning Systems. R. Krosley, M. Misra, Symbolic Distributed Representations. Part V: Epilog F. Alexandre, An Analysis of Connectionist-Symbolic Integration.

Artificial neural networks (ANNs) are widely used for classification tasks where discriminant cues and also contextual parameters are proposed as ANN inputs. When the input space is too large to enable a robust, time limited learning, a classical solution consists in designing a set of ANNs for different context domains. We have proposed a new learning algorithm, the lateral contribution learning algorithm (LCLA), based on the backpropagation learning algorithm, which allows for such a solution with a reduced learning time and more efficient performances thanks to lateral influences between networks. This attractive, but heavy solution has been improved thanks to the orthogonal weight estimator (OWE) technique, an original architectural technique which, under light constraints, merges the set of ANNs in one ANN whose weights are dynamically estimated, for each example, by others ANNs, fed by the context. This architecture allows to give a very rich and interesting interpretation of the weight landscape. We illustrate this interpretation with two examples: a mathematical function estimation and a process modelization used in neurocontrol.

Studies about presetting temper mills have traditionally used mathematical tools. But artificial neural networks can help solve the problem of modeling complex functions in high-dimensional spaces. On-line implementation of an ANN at a temper mill plant has led to a hybrid model that mixes symbolic and connectionist modules.

International audience; The orientation of sensors toward regions of interest of the environment is an important motor activity, monitored by ancient struc- tures of the brainstem. Particularly, the superior colliculus is known to be deeply involved in visual saccadic behavior. Target selection relies on various hints including exogenous information about the nature and the position of candidate targets and endogenous information about current motivations. We present a model of the collicular structure based on bio- logical data, the speci city of which is related to the homogeneity of the underlying substratum of computation. This makes it more suitable to process massive visual ows on a distributed architecture, as it could be requested in a realistic task in autonomous robotics. The present model is restricted to the exogenous part of the visual pathway, from the retina to the superior colliculus. A realistic behavior for the selection of exogenous targets is reported here

We propose an on-line learning algorithm for Multi Layered Perceptrons (MLP) with an Orthogonal Weight Estimator (OWE) architecture. Such an architecture allows to dynamically and efficiently estimate the weights of a MLP in context dependent behaviour problems. The proposed learning algorithm attempts to solve the problem of time-consuming in the learning phase encountered to train these weight estimators.

Basal Ganglia, a group of sub-cortical neuronal nuclei in the brain, are commonly described as the neuronal substratum to Reinforcement Learning. Since the seminal work by Schultz [1], a huge amount of work has been done to deepen that analogy, from functional and anatomic points of view. Nevertheless, a noteworthy architectural hint has been hardly explored: the outstanding reduction of dimensionality from the input to the output of the basal ganglia. Bar-Gad et al. [2] have suggested that this transformation could correspond to a Principal Component Analysis but did not explore the full functional consequences of this hypothesis. In this paper, we propose to study this mechanism within a model more realistic from a computational neuroscience point of view. Particularly, we show its feasibility when the loop is closed, in the framework of Action Selection.

International audience; In this paper, dynamic neural fields are used to develop key features of a cortically-inspired computational module. Under the perspective of designing computational systems that can exhibit the flexibility and genericity of the cortical substrate, using neural field as the competition layer for self-organizing modules has to be considered. However, despite the fact that they serve as a biologically-inspired model, applying dynamic neural fields to drive self-organization is not straightforward. In order to address that issue, an original method for evaluating neural field equations is proposed, based on statistical measurements of the field behavior in some scenarios. Limitations of classical neural field equations are then quantified, and an original field equation is proposed to overcome these difficulties. The performance of the proposed field model is discussed in comparison with some previously considered models, leading to the promotion of the proposed model as a suitable mean for processing competition in cortex-like computation for cognitive systems.

International audience; By applying a modulatory feedback upon the BCM learning rule, we show that it is possible to introduce a bias in the learning process, thus promoting a certain kind of selectivity on the feedforward path.

We propose in this paper a new connectionist unit that matches a biological model of the cortical column. The architectural and functional characteristics of this unit have been designed in the simplest manner in order to simulate human-like reasoning, and to be as similar as possible to the main known features of real intracortical networks. We use a new type of learning rule which can easily take into account goal-oriented combinations of actions in behavioral programs. These learning rules are both simple and biologically plausible. We show in this paper that such units can be used in multilayered networks to perform pattern recognition, with feedback connections effecting an attentive gating of sensory information flow. Computer simulations were performed to assess the ability of a multilayered network made of these biologically inspired units to perform standard speech and visual recognition. Such simulations show levels of performance equivalent to the best currently available connectionist networks for typical human-like problems, with very fast learning and recognition processes. Furthermore, this type of “cortical” unit can be used in more general multilayered networks with units controlling different types of external processing, in order to learn programs of actions which may be included in the process of recognition.

Understanding the brain goes through the assimilation of an increasing amount of biological data going from single cell recording to brain imaging studies and behavioral analysis. The description of cognition at these three levels provides us with a grid of analysis that can be exploited for the design of computational models. Beyond data related to specific tasks to be emulated by models, each of these levels also lays emphasis on principles of computation that must be obeyed to really implement biologically inspired computations. Similarly, the advantages of such a joint approach are twofold: computational models are a powerful tool to experiment brain theories and assess them on the implementation of realistic tasks, such as visual search tasks. They are also a way to explore and exploit an original formalism of asynchronous, distributed and adaptive computations with such precious properties as self-organization, emergence, robustness and more generally abilities to cope with an intelligent interaction with the world. In this article, we first discuss three levels at which a cortical circuit might be observed to provide a modeler with sufficient information to design a computational model and illustrate this principle with an application to the control of visual attention.

International audience; Although biomimetic autonomous robotics relies on the massively parallel architecture of the brain, a key issue for designers is to temporally organize behaviour. The distributed representation of the sensory information has to be coherently processed to generate relevant actions. In the visuomotor domain, we propose here a model of visual exploration of a scene by the means of localized computations in neural populations whose architecture allows the emergence of a coherent behaviour of sequential scanning of salient stimuli. It has been implemented on a real robotic platform exploring a moving and noisy scene including several identical targets.

This paper presents a computational self-organizing model of multi-modal information, inspired from cortical maps. It shows how the organization in a map can be influenced by the same process occurring in other maps. We illustrate this approach on a phonetic - motor association, that shows that the organization of words can integrate motor constraints, as observed in humans.

The main aim in this paper is to study an algorithm of vigilance detection from a minimal number of EEG electrodes, easy to implement on programmable devices, to be used in ambulatory and real everyday life conditions. The connectionist unsupervised approach is summarized in this paper. From the unsupervised classification obtained, a connectionist supervised classification algorithm, the learning vector quantization (LVQ), is used for two different tasks. Firstly, the artefacted states are detected and removed. Secondly, the states deprived of artefacts are then classified in order to decide for the state of vigilance. Connectionist methods with supervised and unsupervised training were used to discriminate the EEG signals characterizing the vigilance states. An artificial neuronal model with a minimal architecture minimizes the complexity and allows implementation. It demonstrates that information, pertinent enough to characterize vigilance states, can be extracted from EEG signal recorded from a single electrode It should also be noted that the intervention of the expert is fundamental in this approach to differentiate nonartefacted vigilance states and artefacted vigilance states.

Colloque avec actes et comité de lecture. internationale.; International audience; This article presents a study performed in the scope of an industrial research project supported by the ECSC. Unsupervised methods of signal encoding and unsupervised extraction of evolution pattern are presented. The main advantages are the ability of proposed clustering method to process arbitrary long input vectors and the adapted coding method that is based on feature extraction of the analysed signals. First results obtained on electric arc furnace data are shown || Cet article présente une étude réalisée dans le cadre d'un projet industriel de recherche subventionné par la CECA. Des méthodes non-supervisées de codage et d'extraction de prototypes d'évolution sont présentées. Les principaux avantages sont l'habilité

Colloque avec actes et comité de lecture. internationale.; International audience; This paper presents an area-saving parallel implementation of a Self-Organizing Map neural network (SOM) on FPGA. The purpose is to make available a finer grain of parallelism to be used in massively SIMD parallel SOM system architectures. We have handled a serial arithmetics (Most Significant Bit First: MSBF and Least Significant Bit First: LSBF), to process the different mathematical operations. Above all, our work has been oriented in such a way to get a light, easy to wear system for classification of vigilance states in humans from electroencephalographic (EEG) signals. The performances of our implementation in terms of area, speed and especially power consumption are highly satisfactory.

Colloque avec actes et comité de lecture. internationale.; International audience; This paper presents three pruning algorithms based on Optimal Brain Surgeon (OBS) and Unit-Optimal Brain Surgeon (Unit-OBS). The first variant performs a backward selection by successively removing single weights from the input variables to the hidden units in a fully connected multilayer perceptron (MLP) for variable selection. The second one removes a subset of non-significant weights in one step. The last one combines the two properties presented above. Simulation results obtained on the Monk's problem illustrate the specificities of each method described in this paper according to the preserved variables and the preserved weights.

Contribution à un ouvrage.; Cognitive modeling can be derived from expert knowledge integration or experimental data analysis. This chapter introduces these two main ways of modeling and concentrates on the latter aspect and the related numerical tools together with their possible role. More precisely, the evolutionary approach is presented for optimization and tuning, artificial neural networks for associative reasoning and the stochastic approach for planning. We then conclude on the need for a modular framework, integrating these numerical and other symbolic tools.

Colloque avec actes et comité de lecture.; Most models concerned with real-world applications can be improved in structuring data and incorporating knowledge about the domain. In our problem of radio electrical wave dying down prediction for mobile communication, a geographic database can be divided in contextual subsets, each representing an homogeneous domain where a predictive model performs better. More precisely, by clustering the input space, a predictive model (here a multilayer perceptron) can be trained on each subspace. Various unsupervised algorithms for clustering were evaluated (Kohonen's maps. Desieno's algorithm, Neural gas, Growing Neural Gas, Buhmann's algorithm) to obtain class homogeneous enough to decrease the predictive error of the radio electrical wave prediction

The authors propose a novel approach to neuronlike models that consists in simulating cortical columns, i.e. associations of neurons having a specific, functional activity. This model is built according to a theory which is consistent with neurobiological data. The model is implemented through a network of columns that simulate the inherently parallel functioning of the nervous system. A description is given of this model, and the experimental conditions in which it is now being tested are described. >

We propose a new architectural model to deal with applications where the input space is very complex and high-dimensional. This solution is inspired from a classical connectionist approach and avoids saturation (A. Guez et al., 1988) due to the use of another connectionist structure, for memory storage. This model is very useful for control problems where input spaces are often complex and heterogeneous. >

The purpose of this work is to deal with real world applications where the size of the problem and the influence of contextual parameters can lead to inefficiency of control including neuro-control. Indeed, the size of real world problems is often very high, especially if numerous contextual parameters are used. This often leads to a prohibitive size for neural networks taken in their monolithic version. On the other hand, the absence of this context will destinate the work to a specific domain of application. These remarks are all the more accurate as we know that this contextual parameters generally have a continuous effect on the process and that they could be considered as specific modulators of the application. We present here a learning algorithm for neuro-controllers with reduced number of inputs. In fact, the contextual parameters are only used to estimate the synaptic weights of the neuro-controller with other NNs. Moreover, the present version is an online economic solution to this problem. >

Symbolic and connectionist tools are widely used in the artificial intelligence (AI) domain. Both approaches report weaknesses to address the full range of capabilities required by this domain, and are often combined to overcome some limitations, but no extensive efforts have been done to give a wider view of the problem. More precisely, many ad hoc hybrid models were built from symbolic and connectionist tools, but little was said about the choice, integration and combination of such tools. Such was the goal of the European ESPRIT project MIX. We report the main conclusions of this project, together with its results on a real-world application.

In this paper, a semi-empirical model of field strength prediction combining theoretical results of propagation loss algorithms and artificial neural networks is considered. This approach expects to overcome some limitations inherent in existing semi-empirical models: linear behaviour of the statistical analysis used in the construction of the models, unfitness for learning new situations. The good results obtained in a dense urban area show that neural networks are a very efficient empirical method to compute new kinds of models which integrate theoretical and experimental data.

Endowing an autonomous system like a robot with intelligent behavior is a difficult problem for several reasons. First, behavior is such a wide topic that a general framework paradigm of inspiration must be chosen in order to obtain a consistent model. Such a framework can be for example biological modeling or an artificial intelligence approach. Second, a general framework is not sufficient to determine a fully specified program to be implemented in a robot. Many choices, tuning and tests must be carried out before obtaining a robust system. This paper describes this. First, a biological model is presented, based on the definition of cortex-like automata, representing elementary functions in the perceptive, motor or associative domain. These automata are connected in a network whose architecture, functioning and learning rules are described in a cortical framework. Second, the computational model derived from that biological model is specified. The way units exchange and compute variables through links is explained, with reference to corresponding biological elements. It is then easier to report experiments allowing an autonomous system to learn regularities of a simple environment and to exploit them to satisfy some internal drives. Even if additional biological hints can be added, this model allow us to better understand how a biological model can be implemented and how biological properties can emerge from a distributed set of units.

Contribution à un ouvrage.; Whereas classical connectionist models can hardly cope with difficult dynamic tasks with a strong temporal factor, many temporal mechanisms inspired with neurobiological data has been proposed in the past and yield efficient time processing properties. The goal of this chapter is to show that, beyond these isolated mechanisms, their integration in a more general architectural and functional framework can potentiate their power and make them usable for non trivial behavioral tasks. We propose a cerebral framework, from the neuronal to the behavioral level, and give some applicative illustrations that underline the encouraging results obtained today.