Search

International audience; We propose a global framework for modeling the cortico-basal system (CX-BG) and the fronto-striatal system (PFC-BG) for the generation and recall of audio memory sequences; ie, sound perception and speech production. Our genuine model is based on the neural architecture called INFERNO standing for Iterative Free-Energy Optimization of Recurrent Neural Networks. Free-energy (FE) corresponds to the prediction error on internal or external noise. FE minimization is used for exploring, selecting and learning in PFC the optimal choices of actions to perform in the BG network (eg sound production) in order to reproduce and control the most accurately possible the spike trains representing sounds in CX. The difference between the two working memories relies in the neural coding itself, which is based on temporal ordering in the CX-BG networks (Spike Timing-Dependent Plasticity) and on the rank ordering in the sequence in the PFC-BG networks (gating or gain-modulation). We detail in this short paper the CX-BG system responsible to encode the audio primitives at few milliseconds order, and the PFC-BG system responsible for the learning of temporal structure in sequences. Two experiments done with a small and a big audio database show the capabilities of exploration, generalization and robustness to noise of the neural architecture to retrieve audio primitives as well as long-range sequences based on structure detection. Although both learning mechanisms are implemented with the same algorithm of rank-order coding, the CX-BG system realizes a model-free recurrent neural network (INFERNO) and the PFC-BG system implements a gated recurrent neural network (INFERNO GATE).

In order to keep trace of information, the brain has to resolve the problem of where information is and how to index new ones. We propose that the neural mechanism used by the prefrontal cortex (PFC) to detect structure in temporal sequences, based on the temporal order of incoming information, has served as second purpose to the spatial ordering and indexing of brain networks. We call this process, apparent to the manipulation of neural 'addresses' to organize the brain's own network, the 'digitalization' of information. Such tool is important for information processing and preservation, but also for memory formation and retrieval.

Contains fulltext : 220410.pdf (Publisher’s version ) (Open Access) 4 p.

We present a robot head with facial gestures, audio and vision capabilities toward the emergence of infant-like social features. For this, we propose a neural architecture that integrates these three modalities following a developmental stage with social interaction with a caregiver. During dyadic interaction with the experimenter, the robot learns to categorize audio-speech gestures of vowels /a/, /i/, /o/ as a baby would do it, by linking someone-else facial expressions to its own movements. We show that multimodal integration in the neural network is more robust than unimodal learning so that it compensates erroneous or noisy information coming from each modality. Therefore, facial mimicry with a partner can be reproduced using redundant audiovisual signals or noisy information from one modality only. Statistical experiments on 24 naive participants show the robustness of our algorithm during human-robot interactions in public environment where many people move and talk all the time. We then discuss our model in the light of human-robot communication, the development of social skills and language in infants.

We present an auto-encoder version of gated networks for learning visuomotor transformations for reaching targets and representating the location of the robot arm. Gated networks use multiplicative neurons to bind correlated images from each others and to learn their relative changes. Using the encoder network, motor neurons categorize the induced visual displacements of the robot arm when applying their corresponding motor commands. Using the decoder network, it is possible to infer back the visual motion and location of the robot arm from the activity of the motor units, aka body image. Using both networks as the same time, near targets can simulate a fictious visual displacement of the robot arm and induce the activation of the most probable motor command for tracking it. Results show the effectiveness of our approach for 2 d-of and 3 do -f robot arms. We discuss then about the network and its use for reaching task and body representation, future works and its relevance for modeling the so-called gain-field neurons in the parieto-motor cortices for learning visuomotor transformation.

We propose a unified framework for modeling the cortico-basal system (CX-BG) and the fronto-striatal system (PFC-BG) for the generation and recall of audio memory sequences; ie, sound perception and speech production. Our genuine model is based on the neural architecture called INFERNO standing for Iterative Free-Energy Optimization of Recurrent Neural Networks. Free-energy (noise) minimization is used for exploring, selecting and learning in PFC the optimal choices of actions to perform in the BG network (eg sound production) in order to reproduce and control the most accurately possible the spike trains representing sounds in CX. The difference between the two working memories relies in the neural coding itself, which is based on temporal ordering in the CX-BG networks (Spike Timing-Dependent Plasticity) and on the rank ordering in the sequence in the PFC-BG networks (gating or gain-modulation). We detail in this paper only the CX-BG system responsible to encode the audio primitives at few milliseconds order, while the PFC-BG system responsible for the learning of temporal structure in sequences will be presented in a complementary paper. Two experiments done with a small and a big audio database show the capabilities of exploration, generalization and robustness to noise of the neural architecture to retrieve audio primitives.

International audience; This paper presents a sensory-motor architecture based on a neural network allowing a robot to recognize vowels in a multi-modal way thanks to human mimicking. The robot autonomously learns to associate its internal state to a human's vowel as an infant would to recognize vowel, and learn to associate congruent information.

International audience; Intelligent tutoring systems are increasingly effective for helping the teacher's work with children. However, these technologies are still poorly used for cognitively impaired infants who display autistic spectrum disorders and intellectual disabilities as they don't adapt easily to each infant. We propose an adaptive learning system called SPEAKY for assisting the learning of lexicon to children with the help of the tutor. SPEAKY present a pair of images and questions of gradual difficulty to each infant and adapt the set of images and questions with respect to the child response. Depending on how their tutors scored the child's response, SPEAKY modifies its model of the learner. We proposed an approach based on the reinforcement learning in order to adapt exercises' difficulty to the level and profile of one infant. Our database contains more than 300 images and we have asked more than 2000 questions in three weeks considering all the exercise sessions. The results confirmed that generalization is not possible and that adaptiveness is important as we found that difficulty is child-specific. Through the results we gathered, we also determine the difficulties and facilities points of each child.

International audience; Human working memory is capable to generate dynamically robust and flexible neuronal sequences for action planning, problem solving and decision making. However, current neurocomputational models of working memory find hard to achieve these capabilities since intrinsic noise is difficult to stabilize over time and destroys global synchrony. As part of the principle of free-energy minimization proposed by Karl Fris-ton, we propose a novel neural architecture to optimize the free-energy inherent to spiking recurrent neural networks to regulate their activity. We show for the first time that it is possible to stabilize iteratively the long-range control of a recurrent spiking neurons network over long sequences. We identify our architecture as the working memory composed by the Basal Ganglia and the Intra-Parietal Lobe for action selection and we make some comparisons with other networks such as deep neural networks and neural Turing machines. We name our architecture INFERNO for Iterative Free-Energy Optimization for Recurrent Neural Network.

International audience; We present a neural architecture capable to control syn-ergistically a flexible robotic model of the human vertebral column toward balance and upward posture. The neural controller is composed of non-linear oscillators that control each vertebre of the column constructed on the principle of tensegrity. They play the role of the central pattern generators in the spinal cords to generate rhythmical patterns and to be entrained to the resonant modes of the tensegrity system. After exploration of the different coordination regimes for different coupling parameters, the top-down controller is able to dynamically select, amplify or inhibit each motor synergy for upward postural balance even with respect to external perturbations. 2 Introduction Animal's musculo-skeletal system is based on a complex network of muscles, bones, nerves, tissues and soft-bodies, which are hard to replicate accurately in robots and to control. The animal's biomechanics are however well-ordered so that the neural control done in the spinal cords can fit its organization and generate the dynamical grouping of the muscles for compliant motion. The hierarchical organization of the motor control a.k.a. the motor synergies has been suggested to diminish the curse of dimensionality of control as enounced by Bernstein. Therefore, the design principles of both the body structure and of the neural model are complementary and have to be considered jointly in order to generate complex motion dynamics. Considering the body, the musculo-skeletal system is always soft and elastic and positioned in a stable or neutral posture. This property is specific to tensile structures, which most biological systems possess as attribute. Eventhough the redundancy and nonlinearity within such dynamical system might be considered as an obstacle for control, the symmetries of the overall structure and the many resonant modes generated can serve to reduce the dimensionality of the control problem. For instance, the control and discovery of the motor synergies may be easier to find by applying synchronization and resonance to these resonant modes. In previous works, we presented a framework based on feedback resonance of chaotic controllers to excite the passive dynam-Figure 1: Vertebral column robot. This tensegrity-based robot is compliant and lightweight , mounted with springs in opposition. The motors are mounted in pairs to produce co-contractions. An IMU is placed at the top of the structure for feedback. ics of several robotic devices and to tune them dynamically to their resonant modes. The control was done indirectly through the coupling parameters between the robot and the oscillators per se. By doing so, we decreased significantly the dimensionality of the control space. We suggested that the coupling parameters are playing the role of the neuro-modulators in the spinal cords, which dynamically trigger the different motor synergies [1]. As our aim is of integrating the structure and the control , we expand our framework to the control of a dorsal spine robot based on tensegrity, see Fig. 1. The homepage of the project is given at 1 with videos of the vertebral column in different configurations. After exploration and cat-egorization of its resonant modes and of its behavioral patterns within the parameters space for different coupling values , we control the spinal cord robot to return back to its rhythmical modes or to its postural configuration based on external feedback perturbations received.

Systematic biases have been found when matching haptic and visual locations. In this demo we can show two things; first we can show that these biases are similar when pointing at a visual target with the index finger or with a handle in a power grip. Second, we show that intermodal biases are not simply the result of a mismatch between the senses and that the transformations of the position information between modalities (and hands) are not simply reversible.

International audience; Infants' ability to orient their actions in space improves dramatically after their sixth month when they start to plan the correct motion of their hands for reaching objects. Recent developmental studies speculate that this enhancement of spatial memory corresponds to the activation of the hippocampal system that shapes the parieto-motor cortices for long-term spatial representation. We suggest that the mechanism of phase precession, which plays an active role in the para-hippocampal cortices to transform the continuous body signals into a precise temporal code could contribute as a key component for learning body-place associations. In a computer simulation of a nine-months old baby, we show how the hippocampal system transforms the input signals from the arm muscles into a phase code, the parietal system uses it then to build a topological map of reaching locations cells combined with eyes vision cells. It follows that one body posture can be retrieved back from estimating visually its location (e.g., for a reaching task).

In this paper, we present a neural architecture aimed to reproduce the qualitative properties of the mirror neurons system which encodes neural representations of actions either performed or observed. Several biological researches have emphasized some of its important aspects, for instance, the tight coupling between the sensorimotor maps, the crucial role of timing (temporal information for encoding and detection), or the neurons connectivity. We attempt to model these in a network of spiking neurons to learn the accurate temporal relationships between the sensorimotor maps. After the learning, the neural connectivity efficiently induces functional capabilities in the whole network exhibiting statistics comparable to small-world networks (e.g., scale-free dynamics and hierarchical organization) similar with observed evidences in the mirror neurons system.

International audience; Embodied action representation and action understanding are the first steps to understand what it means to communicate. We present a biologically plausible mechanism to the representation and the recognition of actions in a neural network with spiking neurons based on the learning mechanism of spike-timing-dependent plasticity (STDP). We show how grasping is represented through the multi-modal integration between the vision and tactile maps across multiple temporal scales. The network evolves into a small-world organization with scale-free dynamics promoting efficient inter-modal binding of the neural assemblies with accurate timing. Finally, it acquires the qualitative properties of the mirror neuron system to trigger an observed action performed by someone else

International audience; In the study of complex systems a fundamental issue is the mapping of the networks of interaction between constituent subsystems of a complex system or between multiple complex systems. Such networks define the web of dependencies and patterns of continuous and dynamic coupling between the system's elements char- acterized by directed flow of information spanning multiple spatial and temporal scales. Here, we propose a wavelet-based extension of transfer entropy to measure directional transfer of information between coupled systems at multiple time scales and demonstrate its effectiveness by studying ͑(a) three artificial maps, (b) physiological recordings, and (c) the time series recorded from a chaos-controlled simulated robot. Limitations and potential extensions of the proposed method are discussed.

We present a framework based on iterative free-energy optimization with spiking neural network for modeling the fronto-striatal system (PFC-BG) for the generation and recall of audio memory sequences. In line with neuroimaging studies done in the PFC, we propose a genuine coding strategy using the gain-modulation mechanism to represent abstract sequences based on the rank and location of items within them only. Based on this mechanism, we show that we can construct a repertoire of neurons sensitive to the temporal structure in sequences from which we can represent any novel sequences. The free-energy optimization is used then to explore and to retrieve the missing indices of the items in the correct order for executive control and compositionality. We show that the gain-modulation permits the network to be robust to variabilities and to have long-term dependencies as it implements a gated recurrent neural network. This model, called Inferno Gate, is an extension of the neural architecture INFERNO standing for Iterative Free-Energy Optimization of Recurrent Neural Networks with Gating or Gain-modulation. In experiments done with an audio database of ten thousand MFCC vectors, Inferno Gate is capable to encode efficiently and retrieve chunks of fifty items length. We discuss then about the potential of our network to model the features of the working memory in PFC-BG loop for structural learning, goal-direction and hierarchical reinforcement learning.