Research Engineer

Le descriptif de l'offre ci-dessous est en Anglais

Type de contrat : CDD

Niveau de diplôme exigé : Bac + 5 ou équivalent

Fonction : Ingénieur scientifique contractuel

A propos du centre ou de la direction fonctionnelle

The Inria center at the University of Bordeaux is one of the nine Inria centers in France and has about twenty research teams.. The Inria centre is a major and recognized player in the field of digital sciences. It is at the heart of a rich R&D and innovation ecosystem: highly innovative SMEs, large industrial groups, competitiveness clusters, research and higher education players, laboratories of excellence, technological research institute...

Contexte et atouts du poste

When we listen to a song, or listening at the radio, our brain needs to parse incoming stimuli incrementally and on the fly. When we learn a song, we learn to imitate what we hear by trial and error, we try to reproduce the sounds we hear. There is converging evidence that (song, language or gesture) production and perception are not separated processes in the brain, they are rather interwoven. This interweaving is for instance what enables people to predict themselves and each other [6]. Interweaving of action and perception is important because it allows a learning agent (e.g. a baby, a bird or a model) to learn from its own actions: for instance, by learning the perceptual consequences (e.g. the heard sounds) of its own actions (e.g. vocal productions) during babbling. Thus, the agent will learn in a self-supervised way. This kind of learning is more biologically plausible than supervised learning which assumes the availability of "teacher signals" which have to be designed by the modeller. Self-supervised learning is fundamental for developmental processes such as babbling. Schwartz et al. [11] propose that perception and action are co-structured in the course of speech development: gestures are perceptually-shaped, they form a perceptuo-motor unit. A clear neuronal model explaining which are the mechanisms shaping such perceptuo-motor units through development is missing.

To learn songs, we need to have a good cognitive representation of sounds and musicality. In order to obtain plausible brain representations of songs or complex sequence of movements we cannot rely on engineered representations (e.g. word or audio embeddings such as Word2Vec or Wave2vec), because this would prevent from modelling the representations obtained during developmental and bootstrapping processes. Thus, we want to obtain perceptuo-motor representations that emerge through action-perception mechanisms. The existence of sensorimotor (i.e. mirror) neurons at abstract representation levels (called action-perception circuits [5]), jointly with the perceptuo-motor shaping of sensorimotor gestures, suggest the existence of similar action-perception mechanisms implemented at different levels of hierarchy (e.g. phoneme, syllable or word in the case of human language). Consequently, models of action-perception mechanisms should be able to be stacked as hierarchical processes.

If we want to recognize a song or to understand a sentence on the fly, our brain needs to process the information as quickly as possible in order to not saturate our "cognitive buffer" (i.e. working memory), thus loosing what is coming next. Christiansen & Chater propose that when the brain is processing a stimulus (e.g. an utterance) it must avoid getting stuck in the "Now or Never Bottleneck" [1]: the brain is forced to extract the necessary information as soon as possible, otherwise the information will be lost. Thus, the rich perceptual input needs to be recoded as it arrives in order to capture the key elements of the sensory information [1]. These compressed (or "chunked") representations are abstractions of inputs (filtering out the details) rather than predictions encoding all the fluctuations of fast incoming inputs. Memory limitations also apply to these recoded representations; hence the brain needs to chunk the compressed representations into multiple levels of representation of increasing abstraction in perception, and decreasing levels of abstraction in action [1]. Therefore, each sequence of chunks at one level will be encoded as a single chunk to a higher level.

Mission confiée

This project will be conducted over a 12-month period, potentially renewable, to allow for in-depth investigation of developmental sequence learning mechanisms.

The general aim of the ANR DeepPool project is to build a dynamic neuronal model of vocal processing and production: the model should be developmental, hierarchical and use action-perception mechanisms. This multi-scale model will span from sensorimotor vocal imitation towards processing and production of long sequences. It will use incremental learning schemes, with goal-directed exploration and seek symbol emergence. We want to create a generic action-perception mechanism that (i) would enable action and perception to shape one another, (ii) while allowing to bootstrap the development of representations from raw sound perceptions, and (iii) which could be stacked as layers of a hierarchical architecture. More info on the ANR DeepPool project:

The engineer project will explore one or several of the topics of the ANR project above. The methods developed will be based on Recurrent Neural Network (RNN), reservoir in particular, but could also use emerging hybrid models in-between Transformers and reservoirs [14] that we create in the team. A reservoir [3] is a random recurrent neural network made of non-linear units that have been used to model various cortical areas [2, 12]. Reservoirs do not involve unfolding of time like BPTT used in LSTMs. In order to build action-perception mechanisms we will embed various concepts from incremental, developmental, reinforcement and unsupervised learning. In particular, we will build on top of preliminary results we have on distal learning with reservoirs [4]. We will also use and develop new reinforcement learning rules adapted to reservoir computing, such as Hebbian exploratory rules [7], that we will combine with unsupervised learning rules that we previously developed such as Dynamic Self-Organizing Maps (DSOM) [9]. Moreover, we will enhance such models with a robust long-term memory mechanism that we recently developed [12].

We will start by implementing the full sensorimotor architecture that we defined in our review [8]. We will build on our recent results both on human speech and birdsong data. For instance, on the songbird side, we built a simple sensorimotor model using a reservoir as the perceptive decoder, a simple Hebbian learning rule for the inverse model, and a Generative Adversarial Network (GAN) as the sound generator given the motor commands. This model is able to reproduce faithfully canary syllables using only 3-dimensional latent space [13, 14]. In order to create the core action-perception layer, the first steps will be to incorporate a forward model and replace the GAN by a reservoir. Later on, we will stack several of these layers at different levels of hierarchy in order to extract chunks (i.e. groups of acoustic elements) of increasing size and complexity. The models will be bootstrapped from goal-directed learning (e.g. vocal imitation). Model features will not to be predefined by the modeller but they will emerge through developmental processes. Because we will be using similar model components, we will be able to apply similar analysis methods, thus facilitating multi-scale analyses.

The RNN mechanisms developed will be applied on human speech and bird songs, because both share similar properties adequate for the project: humans and birds learn to imitate the complex sounds that their fellows produce; they developmentally learn them starting from a babbling exploration phase; both bird songs and human language share a hierarchical organisation of elements with increasing chunk sizes; temporal context is key to make decisions on chunks (i.e. delimitation of chunk boundaries is ambiguous if the context is ignored); and vocal production models are available for both human and bird (e.g. VocalTractLab for human voice) [8].

Generic models, such as random reservoirs, can have a cross-domain impact, opening potential adaptations to non-vocal tasks. The methods and neural mechanisms that will be developed will not be limited to audio applications, but will be generic enough to be also applied to other domains such as motor gesture learning. Because such methods will be based on online, incremental and loosely supervised learning, they could provide more efficient methods useful for machine learning and artificial intelligence domains. Moreover, such sensorimotor models will be use as tools to analyse neuroscience experimental data of our collaborators with a new perspective, and could help in the long run to better understand mechanisms at work in speech rehabilitation therapies.

[1] M. H. Christiansen, N. Chater, P. W. Culicover. Creating language: Integrating evolution, acquisition, and processing. MIT Press, 2016.

[2] Hinaut, P.F. Dominey. Real-Time Parallel Processing of Grammatical Structure in the Fronto-Striatal System: A Recurrent Network Simulation Study Using Reservoir Computing. PloS ONE 8(2): e

[3] H. Jaeger, H. Haas Harnessing nonlinearity: Predicting chaotic systems and saving energy in wireless communication. science, , 78-80.

[4] Juven, X. Hinaut. Cross-Situational Learning with Reservoir Computing for Language Acquisition Modelling. International Joint Conference on Neural Networks, Glasgow, UK. July 2020.

[5] F. Pulvermüller, L. Fadiga. Active perception: sensorimotor circuits as a cortical basis for language. Nature Reviews Neuroscience, 11(5):351–360, Apr. 2010.

[6] M. Pickering, S. Garrod. An integrated theory of language production and comprehension. Behavioral and brain sciences, 36(4):329–347, 2013.

[7] A. Pitti et al. Gated spiking neural network using iterative free-energy optimization and rank-order coding for structure learning in memory sequences. Neural Networks, 121:242–258, Jan. 2020.

[8] Pagliarini, A. Leblois, and X. Hinaut. Vocal imitation in sensorimotor learning models: a comparative review. IEEE Journal of Transaction in Cognitive Develomental Systems. 2020.

[9] Rougier, & Y. Boniface Dynamic self-organising map. Neurocomputing, 74(11),

[10] Rougier Pourquoi votre chat est nul aux échecs et pourtant plus intelligent qu'une IA. The Conversation. hal

[11] J.-L. Schwartz, A. Basirat, L. Ménard, and M. Sato. The perception-for-action-control theory (PACT): A perceptuomotor theory of speech perception. J. of Neuroling., 25(5):336–354, Sept. 2012.

[12] Strock, X. Hinaut, N. Rougier A Robust Model of Gated Working Memory. Neural Computation, Massachusetts Institute of Technology Press (MIT Press), pp.1-29.

[13] Pagliarini, A. Leblois, and X. Hinaut Canary Sensorimotor Model with RNN-Decoder and Low-dimensional GAN Generator. ICDL

[14] Y. Bendi-Ouis, X. Hinaut (preprint 2025) Echo State Transformer: Attention Over Finite Memories Preprint, hal v2. https://hal.science/hal /

Principales activités

• Developping resevoir models based on ReservoirPy and integrate them in the ReservoirPy github as tools for the community
• Use computer clusters (Plafrim, Jean Zay, ...) to evaluate big version of the model to test how their scale to numerous and/or high-dimensional data
• Supervise interns related to the project
• Collaborate with colleagues to find correlates between models developped and birds/human recordings (e.g. fMRI, electrophysiology recordings)
• Communicate on the results in conferences and journals
• Contribute to current team projects and collaborations with his/her expertise

Compétences

Required Knowledge and background

• Good background in maths and computer science;
• A strong interest for neuroscience and the cognitive processes underlying learning;
• Python programming with experience in scientific libraries Numpy/Scipy (or similar programming language: matlab, etc.);
• Experience in machine learning or data mining is preferred;
• Independence and ability to manage a project;
• Good English reading/writing/speaking skills.

Avantages

• Subsidized meals
• Partial reimbursement of public transport costs
• Leave: 7 weeks of annual leave + 10 extra days off due to RTT (statutory reduction in working hours) + possibility of exceptional leave (sick children, moving home, etc.)
• Possibility of teleworking (after 6 months of employment) and flexible organization of working hours
• Professional equipment available (videoconferencing, loan of computer equipment, etc.)
• Social, cultural and sports events and activities
• Access to vocational training
• Social security coverage

Rémunération

grossly remuneration depending on qualifications and professional experiences :

from 2692 per month to 3085 euros (before taxs)

Informations générales

• Thème/Domaine : Neurosciences et médecine numériques

Biologie et santé, Sciences de la vie et de la terre (BAP A)
- Ville : Bordeaux
- Centre Inria : Centre Inria de l'université de Bordeaux
- Date de prise de fonction souhaitée :
- Durée de contrat : 12 mois
- Date limite pour postuler :

Attention: Les candidatures doivent être déposées en ligne sur le site Inria. Le traitement des candidatures adressées par d'autres canaux n'est pas garanti.

Consignes pour postuler

Thank you to send:

• A detailed CV
• A motivation letter
• Support letters (mandatory)

Sécurité défense :

Ce poste est susceptible d'être affecté dans une zone à régime restrictif (ZRR), telle que définie dans le décret n° relatif à la protection du potentiel scientifique et technique de la nation (PPST). L'autorisation d'accès à une zone est délivrée par le chef d'établissement, après avis ministériel favorable, tel que défini dans l'arrêté du 03 juillet 2012, relatif à la PPST. Un avis ministériel défavorable pour un poste affecté dans une ZRR aurait pour conséquence l'annulation du recrutement.

Politique de recrutement :

Dans le cadre de sa politique diversité, tous les postes Inria sont accessibles aux personnes en situation de handicap.

Contacts

• Équipe Inria : MNEMOSYNE
• Recruteur :

Hinaut Xavier /

A propos d'Inria

Inria est l'institut national de recherche dédié aux sciences et technologies du numérique. Il emploie 2600 personnes. Ses 215 équipes-projets agiles, en général communes avec des partenaires académiques, impliquent plus de 3900 scientifiques pour relever les défis du numérique, souvent à l'interface d'autres disciplines. L'institut fait appel à de nombreux talents dans plus d'une quarantaine de métiers différents. 900 personnels d'appui à la recherche et à l'innovation contribuent à faire émerger et grandir des projets scientifiques ou entrepreneuriaux qui impactent le monde. Inria travaille avec de nombreuses entreprises et a accompagné la création de plus de 200 start-up. L'institut s'efforce ainsi de répondre aux enjeux de la transformation numérique de la science, de la société et de l'économie.

---