Search

We propose a developmental model inspired by the cortico-basal system (CX-BG) for vocal learning in babies and for solving the correspondence mismatch problem they face when they hear unfamiliar voices, with different tones and pitches. This model is based on the neural architecture INFERNO standing for Iterative Free-Energy Optimization of Recurrent Neural Networks. Free-energy minimization is used for rapidly exploring, selecting and learning the optimal choices of actions to perform (eg sound production) in order to reproduce and control as accurately as possible the spike trains representing desired perceptions (eg sound categories). We detail in this paper the CX-BG system responsible for linking causally the sound and motor primitives at the order of a few milliseconds. Two experiments performed with a small and a large audio database show the capabilities of exploration, generalization and robustness to noise of our neural architecture in retrieving audio primitives during vocal learning and during acoustic matching with unheared voices (different genders and tones)., Author summary We designed a developmental architecture inspired by the cortico-basal system for early vocal learning. Our neural system explores, evaluates and strengthens the motor primitives that match the best the sound repertoire created also dynamically. After a babbling process in which the network tests and aligns pronounced sound and motor vocal tracks, it is used for listening to novel voices, solving the correspondence problem.

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.

International audience; We present a framework based on iterative free-energy optimization with spiking neural networks for modeling the fronto-striatal system (PFC-BG) for the generation and recall of audio memory sequences. In line with neuroimaging studies carried out in the PFC, we propose a genuine coding strategy using the gain-modulation mechanism to represent abstract sequences based solely on the rank and location of items within them. 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. Free-energy optimization is then used 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 mechanism 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 performed with an audio database of ten thousand MFCC vectors, Inferno Gate is capable of encoding efficiently and retrieving chunks of fifty items length. We then discuss the potential of our network to model the features of working memory in the PFC-BG loop for structural learning, goal-direction and hierarchical reinforcement learning.

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 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.

Adults readily make associations between stimuli perceived consecutively through different sense modalities, such as shapes and sounds. Researchers have only recently begun to investigate such correspondences in infants but only a handful of studies have focused on infants less than a year old. Are infants able to make cross-sensory correspondences from birth? Do certain correspondences require extensive real-world experience? Some studies have shown that newborns are able to match stimuli perceived in different sense modalities. Yet, the origins and mechanisms underlying these abilities are unclear. The present paper explores these questions and reviews some hypotheses on the emergence and early development of cross-sensory associations and their possible links with language development. Indeed, if infants can perceive cross-sensory correspondences between events that share certain features but are not strictly contingent or co-located, one may posit that they are using a “sixth sense” in Aristotle’s sense of the term. And a likely candidate for explaining this mechanism, as Aristotle suggested, is movement.

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.

In this paper we explore the dynamics of Joint Attention (JA) in children with Autism Spectrum Disorder (ASD) during an interaction task with a small humanoid robot. While this robot elicits JA in children, a coupled perception system based on RGB-D sensors is able to capture their behaviours. The proposed system shows the feasibility and the practical benefits of the use of social robots as assessment tools of ASD. We propose a set of measures to describe the behaviour of the children in terms of body and head movements, gazing magnitude, gazing directions (left vs. front vs. right) and kinetic energies. We assessed these metrics by comparing 42 children with ASD and 16 children with typical development (TD) during the JA task with the robot, highlighting significant differences between the two groups. Employing the same metrics, we also assess a subgroup of 14 children with ASD after 6-month of JA training with a serious game. The longitudinal data confirms the relevance of the proposed metrics as they reveal the improvements of children behaviours after several months of training.

In this work, we study how learning in a special environment such as a museum can influence the behavior of robots. More specifically, we show that online learning based on interaction with people at a museum leads the robots to develop individual preferences. We first developed a humanoid robot (Berenson) that has the ability to head toward its preferred object and to make a facial expression that corresponds to its attitude toward said object. The robot is programmed with a biologically-inspired neural network sensory-motor architecture. This architecture allows Berenson to learn and to evaluate objects. During experiments, museum visitors’ emotional responses to artworks were recorded and used to build a database for training. A similar database was created in the laboratory with laboratory objects. We use those databases to train two simulated populations of robots. Each simulated robot emulates the Berenson sensory-motor architecture. Firstly, the results show the good performance of our architecture in artwork recognition in the museum. Secondly, they demonstrate the effect of training variability on preference diversity. The response of the two populations in a new unknown environment is different; the museum population of robots shows a greater variance in preferences than the population of robots that have been trained only on laboratory objects. The obtained diversity increases the chances of success in an unknown environment and could favor an accidental discovery.

International audience; In this paper we propose a neural network allowing a mobile robot to learn artwork appreciation. The learning is based on the social referencing approach. The robot acquires its knowledge (artificial taste) from the interaction with humans. We present and analyze specifically the visual system, its impact on the robot behavior, and at the end, we analyze the readability of our robot behavior according to visitors comments. We show that the low level spatial competition between the values associated to areas of interest in the image are important for the coherence of the robot's object evaluation and the readability of its behavior. Index Terms— Artificial intelligence, neural networks, Human robot interactions, computer vision.

Resume Le nombre d’etudes sur l’utilisation des technologies de l’information, de la communication et de la robotique au service des enfants autistes a progresse rapidement au cours des 15 dernieres annees. Nous detaillons les espoirs, mais aussi les ecueils et les verrous technologiques des deux domaines les plus dynamiques, a savoir : (1) la realisation de jeux serieux destines a entrainer des habiletes specifiques (reconnaissance des emotions, interactions sociales) ; (2) les interactions avec des robots. Nous prendrons plus particulierement deux exemples a partir de projets auxquels nous collaborons ou avons directement collabore, JeMiMe pour les jeux serieux, et Michelangelo pour la robotique. Nous concluons que si nous sommes convaincus de l’impact potentiel de ces methodes dans la prise en charge des enfants autistes, les limites sont encore importantes et la validite clinique tres limitee.

International audience

We aimed to compare, during a joint attention (JA) elicitation task, how children with autism spectrum disorder (ASD) and children with typical development (TD) behave and explore their 4 dimensional (meaning spatial 3D + time) when interacting with a human or with a robotic agent. We built a system that employed a Nao robot and a perception system based on a RGB-D sensor (Kinect) to capture social engagement cues. A JA induction experiment was performed in which children with ASD ( N = 16) and matched TD children ( N = 16) had a 3-min interaction with the robot or with a therapist. Nao induced JA by gazing; by gazing and pointing; and by gazing, pointing and vocalizing at pictures. Both groups of children performed well with the therapist. However, with Nao, both groups had lower JA scores, and the children with ASD had a significantly lower score than the TD children. We found that (i) multimodal JA induction was more efficient in both groups; (ii) the 3D spatial world gaze exploration showed less accuracy; and (iii) the trunk position in ASD showed less stability in the 4 dimensions compared to TD controls. We conclude that, in ASD, JA skill depends on the interaction partner, and implies a higher motor and cognitive cost.

In this paper, we show that a robotic system can learn online to recognize facial expressions without having a teaching signal associating a facial expression with a given abstract label (e.g., ‘sadness’, ‘happiness’). Moreover, we show that recognizing a face from a non-face can be accomplished autonomously if we imagine that learning to recognize a face occurs after learning to recognize a facial expression, and not the opposite, as it is classically considered. In these experiments, the robot is considered as a baby because we want to understand how the baby can develop some abilities autonomously. We model, test and analyze cognitive abilities through robotic experiments. Our starting point was a mathematical model showing that, if the baby uses a sensory motor architecture for the recognition of a facial expression, then the parents must imitate the baby’s facial expression to allow the online learning. Here, a first series of robotic experiments shows that a simple neural network model can control a robot head and can learn online to recognize the facial expressions of the human partner if he/she imitates the robot’s prototypical facial expressions (the system is not using a model of the face nor a framing system). A second architecture using the rhythm of the interaction first allows a robust learning of the facial expressions without face tracking and next performs the learning involved in face recognition. Our more striking conclusion is that, for infants, learning to recognize a face could be more complex than recognizing a facial expression. Consequently, we emphasize the importance of the emotional resonance as a mechanism to ensure the dynamical coupling between individuals, allowing the learning of increasingly complex tasks.

International audience; In this work, we study how emotional interactions with a social partner can bootstrap increasingly complex behaviors such as social referencing. Our idea is that social referencing as well as facial expression recognition can emerge from a simple sensory-motor system involving emotional stimuli. Without knowing that the other is an agent, the robot is able to learn some complex tasks if the human partner has some "empathy" or at least "resonate" with the robot head (low level emotional resonance). Hence, we advocate the idea that social referencing can be bootstrapped from a simple sensory-motor system not dedicated to social interactions.

Imitation is a powerful paradigm for learning new actions and tasks and is currently investigated in several research domains such as developmental psychology or robotics [8]. For example, learning by demonstration is employed in robotics to teach new skills and tasks to robots [1] [2]. The dynamics of coupling of perception of one partner to the action of the other, and vice-and-versa, has been exploited to characterize social traits [9]. In particular, in the case of strong coupling such as imitation, research in developmental science has shown that imitation could be considered a social referencing mechanism [7]. Recently, identification of social traits during human-robot imitation learning has been successfully demonstrated. For example a social signature emerges from a neural network architecture and distinguishes typical children and children with Autism Spectrum Disorders (ASD). The key idea of this approach is to learn the mapping between the perception of postures performed by the children and the actions performed by the robot. A perception-action architecture based on neural network (PerAc [6]) is used to associate the perception with action [3]. Interestingly, this approach generalizes to person's identity recognition by the analysis of the dynamics of the social signature (novelty detection procedure) [5].

“Mirroring” : A Solution for the Development of Cognitive Skills. In this work, we study the “ mirroring” in the early stages of learning in the context of autonomous system. We propose three experiences that highlight how imitation, facial expressions recognition and social referencing can emerge from a robotic architecture that is not able to distinguish between self and other. Three results are discussed : (1) the “ambiguity” of perception produces the emergence of simple gestural imitation in a system that has previously learned to associate its own perceptions ; (2) facial expressions recognition emerges from a similar architecture, the human plays the role of mirror (social mirroring) through an imitative game ; then by extension ; (3) the development of social referencing incorporates the ability to recognize, understand, respond to and regulate its behavior in response to human’s emotional expression (double mirroring). This work is based on two mechanisms : (1) a sensorimotor architecture based on associative mechanisms allowing to develop a perception as a set of dynamic attractors ; (2) mirroring mechanism allowing the robot to associate what it does with what it sees. We will show that these two assumptions are sufficient to build a sensorimotor architecture that is able to develop cognitive abilities increasingly complex., Dans ce travail, nous étudions la place du mirroring dans les premières étapes de l''apprentissage d''un système autonome. Nous proposons trois expériences qui illustrent comment l’imitation, la reconnaissance des expressions faciales et la référenciation sociale peuvent émerger d''une architecture robotique qui n''est pas capable de faire une différenciation entre soi et l''autre. Trois résultats sont discutés : (1) l’ambiguïté de la perception engendre l''émergence d''imitations gestuelles simples chez un système qui a appris auparavant à associer ses propres perceptions ; (2) la reconnaissance des expressions faciales émerge d''une architecture similaire quand l''humain fait partie du babillage et de la phase d''apprentissage (mirroring social) au travers d’un jeu imitatif ; puis par extension ; (3) le développement de la référenciation sociale incorporant la capacité de reconnaitre, comprendre, répondre et de réguler son comportement en réponse à l’expression émotionnelle du partenaire (double mirroring). Ce travail est basé sur deux mécanismes importants : (1) une architecture sensorimotrice basée sur des mécanismes d’associations permet de développer une perception comme étant un ensemble d’attracteurs dynamiques ; (2) le mécanisme de mirroring permet au robot d’associer ce qu’il fait avec ce qu’il voit. Nous montrerons dans la suite de cet article que ces deux hypothèses sont suffisantes pour construire une architecture sensori-motrice capable de développer des capacités cognitives de plus en plus complexes., Boucenna Sofiane, Andry Pierre, Gaussier Philippe. Le « mirroring » : une solution pour le développement de capacités cognitives. In: Intellectica. Revue de l'Association pour la Recherche Cognitive, n°65, 2016/1. Nouvelles approches en Robotique Cognitive. pp. 81-109.

Classical models of emotions consider either the communicational aspect of emotions (for instance the emotions conveyed by the facial expressions) or the second order control necessary for survival purpose when the autonomy of the system is an issue. Here, we show the interdependence of communication and meta-control aspects of emotion. We propose the idea that emotions must be understood as a dynamical system linking two controllers: one devoted to social interactions (i.e. communication aspects) and another one devoted to the interactions within the physical world (i.e metacontrol of a more classical controller). Illustrations will be provided from applications involving navigation among different goal places according to different internal drives or object grasping and avoidance.

International audience; To interact and cooperate with humans in their daily-life activities, robots should exhibit human-like " intelligence ". This skill will substantially emerge from the interconnection of all the algorithms used to ensure cognitive and interaction capabilities. While new robotics technologies allow us to extend such abilities, their evaluation for social interaction is still challenging. The quality of a human-robot interaction can not be reduced to the evaluation of the employed algorithms: we should integrate the engagement information that naturally arises during interaction in response to the robot's behaviors. In this paper we want to show a practical approach to evaluate the engagement aroused during interactions between humans and social robots. We will introduce a set of metrics useful in direct, face to face scenarios, based on the behaviors analysis of the human partners. We will show how such metrics are useS. M. Anzalone · M. Chetouani Sorbonne Universités , ful to assess how the robot is perceived by humans and how this perception changes according to the behaviors shown by the social robot. We discuss experimental results obtained in two human-interaction studies, with the robots Nao and iCub respectively.

Humans are experts in understanding others' intention. This capacity is essential to interact and collaborate with others. Thus, providing a robot with this capacity is a key improvement for further developments in socially interactive robotics. This work exploits a mirroring scenario between aa caregiver and a robot to develop an intention inference capacity without providing a priori knowledge to the robot. A sensory-motor architecture is used to learn autonomously associations between the robot's internal states and its perception.

In this paper, a robot learns different postures by imitating several partners. We assessed the effect of the type of partners, i.e., adults, typically developing (TD) children and children with autism spectrum disorder (ASD), on robot learning during an imitation game. The experimental protocol was divided into two phases: 1) a learning phase, during which the robot produced a random posture and the partner imitated the robot; and 2) a phase in which the roles were reversed and the robot had to imitate the posture of the human partner. Robot learning was based on a sensory-motor architecture whereby neural networks (N.N.) enabled the robot to associate what it did with what it saw. Several metrics (i.e., annotation, the number of neurons needed to learn, and normalized mutual information) were used to show that the partners affected robot learning. The first result obtained was that learning was easier with adults than with both groups of children, indicating a developmental effect. Second, learning was more complex with children with ASD compared to both adults and TD children. Third, learning with the more complex partner first (i.e., children with ASD) enabled learning to be more easily generalized.

Recently, there have been considerable advances in the research on innovative information communication technology (ICT) for the education of people with autism. This review focuses on two aims: (1) to provide an overview of the recent ICT applications used in the treatment of autism and (2) to focus on the early development of imitation and joint attention in the context of children with autism as well as robotics. There have been a variety of recent ICT applications in autism, which include the use of interactive environments implemented in computers and special input devices, virtual environments, avatars and serious games as well as telerehabilitation. Despite exciting preliminary results, the use of ICT remains limited. Many of the existing ICTs have limited capabilities and performance in actual interactive conditions. Clinically, most ICT proposals have not been validated beyond proof of concept studies. Robotics systems, developed as interactive devices for children with autism, have been used to assess the child's response to robot behaviors; to elicit behaviors that are promoted in the child; to model, teach and practice a skill; and to provide feedback on performance in specific environments (e.g., therapeutic sessions). Based on their importance for both early development and for building autonomous robots that have humanlike abilities, imitation, joint attention and interactive engagement are key issues in the development of assistive robotics for autism and must be the focus of further research.

In the field of biology, the study of bonding has been renewed by the discovery of non genetic transmission of behavioural traits through early mother-infant interaction and the role of stress hormones and ocytocin. However, the study of early interaction is complex and Social Signal Processing (SSP) can help in addressing some issues. Based on works from our group, we will show data from diverse sources (e.g. experiments, home movies) showing how SSP was used to address synchrony between partners (e.g. infant, child, care giver, agent) and characteristics that participates to interpersonal exchanges (e.g. motherese, emotional prosody or faces).

In this paper, we investigate a sensory-motor architecture allowing a robot to learn to recognize postures. The learning is performed without a teaching signal that associates a specific posture with the robot's motor internal state. Our architecture assumes that the robot initially performs postures, then the human imitates them. An on-line learning scheme without an explicit reward or ad-hoc detection mechanism or a formatted teaching technique is proposed. Investigations on how a “naive” system can learn to imitate correctly another person's posture during a natural interaction motivate the current research work.

International audience; Mimicry and deferred imitation have often been considered as separate kinds of imitation. In this paper, we present a simple architecture for robotic arm control which can be used for both. The model is based on some dynamical equations, which provide a motor control with exploring and converging capacities. A visuo-motor map is used to associate positions of the end effector in the visual space with proprioceptive position of the robotic arm. It enables a fast learning of the visuo-motor associations without needing to embed a priori information. The controller can be used both for accurate control and interaction. It has been implemented on a minimal robotic setup showing some interesting emergent properties. The robot can reproduce simple gestures in mimicry situation and finalized actions in deferred imitation situation. Moreover, it can even show some “intention” recognition abilities. Finally, the experiment of deferred imitation which inherits from learning by demonstration, also provides a good basis for cooperative and interactive experiments.

International audience; We are interested in understanding how babies learn to recognize facial expressions without having a teaching signal allowing to associate a facial expression to a given abstract label (i.e the name of the facial expression 'sadness', 'happiness'...). Our starting point was a mathematical model showing that if the baby uses a sensory motor architecture for the recognition of the facial expression then the parents must imitate the baby facial expression to allow the on-line learning. In this paper, a first series of robotics experiments showing that a simple neural network model can control the robot head and learn on-line to recognize the facial expressions (the human partner imitates the robot prototypical facial expressions) is presented. We emphasize the importance of the emotions as a mechanism to ensure the dynamical coupling between individuals allowing to learn more complex tasks.

International audience; A wide variety of visual recognition systems are developed for precise tasks and types of objects. In this paper we would like to emphasize ways to build a more generic recognition system. Perception is one of these mechanisms that psychologists particularly pointed out as a fundamental one for actively organizing and making sense of input sensory information. Based on psychological assumptions, we propose to explore the concept of perception, infer formalization in the dynamical system framework and quantitatively analyze it on robotic platforms using a unique simple neuronal architecture based on the association of visual and motor information (movements of the body or part of the body). This coupling of sensory flows of information can be characterized by a sensorimotor invariant, a dynamical attractor that we identify as a perception function. For place, object or facial expression recognition, we show how simple sensori-motor architecture can be applied to accomplish each task in terms of behavioral recognition. In each application, some pertinent visual information, based on classical focus point detection, are organized as local views and associated to an action or an internal state corresponding to a set of actions, in order to reach a location, an object or recognize a facial expression. The active learning phase for different points of view or face expressions allows the emergence of a stable perception linked to a stable sensori-motor attractor and allows the robot to perform a stable behavior in very different initial conditions. We will show how the attractor/perception emerges during the learning phase and evaluate its spatial generalization properties.

International audience; In this work, we are interested in understanding how emo- tional interactions with a social partner can bootstrap increasingly com- plex behaviors such as social referencing. Our idea is that social refer- encing as well as facial expression recognition can emerge from a simple sensori-motor system involving emotional stimuli. Without knowing that the other is an agent, the robot is able to learn some complex tasks if the human partner has some “empathy” or at least “resonate” with the robot head (low level emotional resonance). Hence we advocate the idea that social referencing can be bootstrapped from a simple sensori-motor system not dedicated to social interactions.

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.