Offre en lien avec l’Action/le Réseau : – — –/– — –

Laboratoire/Entreprise : LIRIS
Durée : 5-6 mois
Contact : stefan.duffner@liris.cnrs.fr
Date limite de publication : 2022-12-31

Contexte :
Deep Neural Networks (DNN) are powerful machine learning models for a large number of applications. However, they may have an enormous number of parameters and require large amounts of memory and computational resources and thus incur a high energy consumption, which makes their use for edge computing difficult.
Several approaches have been proposed to alleviate this problem, e.g. pruning, quantisation or architectural optimisations such as Neural Architecture Search. Although more and more efficient solutions exist also on the practical side (TensorFlow Lite, PyTorch quantization [1], NVIDIA Tensor RT etc.), the deployment of large DNN on embedded systems is still challenging.
Thus, on a more global level, a major concern in reducing the energy consumption related to AI in the cloud as well as on the edge is to make these tools more efficient and more accessible to a larger public.

[1] https://pytorch.org/blog/introduction-to-quantization-on-pytorch/
[2] Renato Cintra, Stefan Duffner, Christophe Garcia & André Leite (2018). « Low-complexity Approximate Convolutional Neural Networks ». IEEE Transactions on Neural Networks and Learning Systems

Sujet :
The goal of this project is to study the state of the art in neural network quantisation and experiment with existing frameworks such as the PyTorch quantisation module. We will particularly focus on post-training static quantisation. The first objective is to implement a simple pipeline (either using one of the existing libraries or from scratch) and make it extensible and adaptable to new algorithms. A set of standard models (MLP and CNN) and some common datasets will serve as a test bench.
A second objective consists in developing and experimenting with new quantisation schemes (fixed-point and floating-point of different precision and different layer-wise/channel-wise strategies).
Finally, a more complex quantisation algorithm that we published earlier [2] should be implemented and adapted to the given framework and pipeline. The developed algorithms should be tested and evaluated (on CPU and GPU hardware).
This internship is part of an industrial exploitation project of research work in collaboration with engineers from Pulsalys (https://www.pulsalys.fr).

Profil du candidat :
Master in Computer Science, AI, machine learning or similar, or final year of engineering school

Formation et compétences requises :
– Good knowledge in machine learning and neural networks
– Knowledge in optimisation is a plus
– Good skills in python programming and Pytorch, scipy, numpy etc.
– Scientific curiosity and creativity
– Motivated to work in a team of researchers and engineers

Adresse d’emploi :
LIRIS – INSA Lyon, 7 Avenue Jean Capelle, 69621 Villeurbanne, France

Offre en lien avec l’Action/le Réseau : – — –/– — –

Laboratoire/Entreprise : LIRIS
Durée : 5-6 mois
Contact : stefan.duffner@liris.cnrs.fr
Date limite de publication : 2022-12-31

Contexte :
While consciousness is currently seen as the result of processes in the brain, ordinary human experience is in fact embedded in a web of causal relations that link the brain to the body and the environment (Bayne et al., 2020). Embodied cognition is a naturalistic theory in which consciousness is associated with a dynamic interaction between brain, body and environment (BBE) (Thompson et Varela, 2001). Indeed, from an evolutionary point of view, the nervous system appears to be dedicated to perceptual and motor processes that allow interaction with the environment (Thompson et Varela, 2001). According to this model, the world of which we are aware is enacted by our past interactions with the environment. In other words, consciousness would be better explained in terms of being in the world at a specific time and place. From a biological point of view, this would be associated with both information shared across cortical regions and non-reflexive behaviors (Kanai et al., 2019).
One way to better understand consciousness is to study its disorders and recovery. Indeed, when such
patients recover, they go through different clinical states that are characterized by the recovery of arousal and/or awareness and by the recovery of BBE interactions. Coma is a state of unconsciousness in which patients cannot be awakened. Those who recover could transit through a disorder of consciousness (DOC).
We hypothesize that the analysis of synchronized brain-behavior recordings to naturalistic stimulations
will better predict consciousness recovery, than the current analyses of one of the two modalities in
neutral contexts. Specifically, we hypothesize that DOC patients who will recover consciousness are those whose cerebral connectivity abilities are associated with non-reflexive behaviors to emotional and/or social stimuli (listening to preferred music, visit from relatives).
The first aim of the project is to apply, in healthy participants and DOC patients, new behavioral tools we have developed thanks to different computer vision methods (body and face analysis from video), and to characterize the behavioral responses associated with naturalistic conditions, thanks to machine-learning methods. The second aim is to characterize, in the same populations, the BBE interactions, i.e. how brain (from high-density EEG connectivity), body (from video and ECG -electrocardiogram) responses are coupled/decoupled according to the environment (emotional and social vs. neutral stimuli). This work will allow a better care of DOC patients and more generally a better understanding of the cerebral-behavioral bases of consciousness.

Sujet :
We will build upon existing preliminary works, where state-of-the-art computer vision algorithms (e.g. OpenPose, OpenFace) for human pose estimation and face and gaze analysis have been integrated and adapted to the specific context and acquisition conditions.
Subjects are successively put into 3 different situations:
• listening to sound,
• rest,
• interactions with the instructor,
and for each situation an emotional and a non-emotional version is performed (e.g. for sound: listening to favorite music). The first experiments based on recurrent (LSTM) neural networks have already given promising results for classifying different situations (in terms of interaction with the environment and overall behavioral phenomena). which we will further improve and make explicit, i.e. interpretable by medical specialists and neuroscientists. For example, at this point, we do not know yet, what parts of the body or face are likely to produce more relevant indicators of consciousness state and which type of behaviour, motion or gestures are informative. This will be one concrete expected result of the internship.
We have video, ECG and hd-EEG data from 20 healthy subjects and 60 DOC patients, which will allow the development of more precise and robust machine learning models.
Nevertheless, this represents relatively few data given the large variability between patients. Thus strong priors and regularisation are needed to avoid overfitting. Also specific pre- and post-processing methods will help to reduce the noise and decrease the dimensionality (e.g. by learning compact feature
embeddings, or by features selection algorithms). For example, we developed a specific filtering algorithm on the output of OpenPose that effectively removes undesired oscillations (due to pixel quantisations) and produces body pose and motion estimates that are much more realistic and that contain fewer artifacts.
Furthermore, different learning strategies and models will be developed to deal with the large amount of noise in general and the imbalance between the amount of relevant data compared to irrelevant data, for example by “rebalancing” the data using other methods or by specific learning mechanisms such as the Multiple Instance learning framework or some type of self-supervised learning.
The combination of these different modalities using new deep learning models as well as the adaptation of our existing models for unsupervised learning multi-variate time series (Berlemont et al. 2017) will allow us to further analyse complex correlations and co-occurences of characteristics and, by focusing on explainable methods and results (explainable AI), give insights into BBE interactions and further give rise to new neuroscientific hypotheses. The findings of such correlation patterns will be another major result that we expect from this project.
Thus, the nature of this research is clearly exploratory, and the expected results will concern both methodological contributions in AI and original methods leading to new knowledge in Neuroscience.

Profil du candidat :
Master in Computer Science, AI, Machine Learning or similar

Formation et compétences requises :
– Good knowledge of machine learning and neural networks and data analysis,
– Strong skills in python programming and PyTorch
– Motivation to work in a multi-disciplinary project
– Scientific curiosity

Adresse d’emploi :
LIRIS – INSA Lyon, 7 Avenue Jean Capelle, 69621 Villeurbanne, France

Offre en lien avec l’Action/le Réseau : – — –/– — –

Laboratoire/Entreprise : CREATIS
Durée : 6 mois
Contact : michael.sdika@creatis.insa-lyon.fr
Date limite de publication : 2023-02-28

Contexte :
Contexte scientifique Dans le cadre d’une collaboration entre les laboratoire Ampère et CREATIS, nous
developpons un simulateur de chirurgie qui permettra l’apprentissage d’un geste à la fois courant et critique:
la ponction ventriculaire. Pour créer des patients virtuels pour le simulateur, nous devons repositionner
l’IRM d’un sujet quelconque sur l’IRM du sujet utilisé pour créer le crâne physique du simulateur (figure
1). Il faudra donc faire correspondre exactement les crânes des deux sujets tout en préservant l’anatomie du
patient virtuel.
Le recalage d’image est un outil permettant d’aligner des images entre elles. Par example sur la figure
2, elle permet de positionner, d’orienter voire de déformer les cerveaux des trois images de droite pour les
replacer dans le système de coordonnées de l’image de gauche. C’est souvent une étape préliminaire essentielle
pour l’étude de pathologie cérébrale basée sur l’imagerie.
Les outils classique de recalage sont souvent basée sur des approches itérative d’optimisation mathématique.
Ces méthodes donnent généralement de bons résultats mais peuvent parfois échouer. Pour palier à ce manque
de robustesse, de plus en plus de méthodes actuelles se basent sur des approches par pprentissage profond
[Boveiri].

Sujet :
Objectifs L’objectif du stage est d’abord de de mettre en place et d’entraı̂ner un réseau de neurones
permettant faire le recalage linéaire d’une image de cerveau sur un espace de référence standard. L’objectif
principale sera que l’estimation soit la plus robuste possible mais aussi que le réseau soit léger. On se basera
sur les résultats obtenus lors d’un stage précédent au cours duquel on se restreignait à des transformations
2D.
On devra ensuite faire la correspondance des crânes par un recalage déformable préservant l’anatomie du
patient. On pourra comparer ici les approches classique et par apprentissage.
Données: plusieurs jeux de données d’imagerie cérébrales publiques impliquant différentes pathologies,
protocoles d’acquisition et modalités sont déjà utilisés dans l’équipe et seront utilisés pour le stage. Une
solide procédure d’augmentation de données permettra d’améliorer encore la robustesse de notre méthode.

Profil du candidat :
Profil du Candidat Le candidat recruté devra avoir une formation dans un des domaines suivants et de
bonnes connaissances dans les deux autres:
• Deep learning
• Traitement d’images
• Mathématiques appliquées
Il devra aussi avoir de solides compétences en développement logiciel en pytorch et être en mesure
d’implémenter les méthodes proposées.
Merci d’envoyer vos candidatures avec CV, lettre de motivation, relevés de notes, lettres de recomman-
dation à michael.sdika[at]creatis.insa-lyon.fr.

Formation et compétences requises :
Profil du Candidat Le candidat recruté devra avoir une formation dans un des domaines suivants et de
bonnes connaissances dans les deux autres:
• Deep learnin)
• Traitement d’images
• Mathématiques appliquées
Il devra aussi avoir de solides compétences en développement logiciel en pytorch et être en mesure
d’implémenter les méthodes proposées.
Merci d’envoyer vos candidatures avec CV, lettre de motivation, relevés de notes, lettres de recomman-
dation à michael.sdika[at]creatis.insa-lyon.fr.

Adresse d’emploi :
Campus de la Doua, Villeurbanne

Document attaché : 202210221945_internship-registration-2023.pdf