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

Laboratoire/Entreprise : Université Grenoble Alpes (UGA), Verimag Laborato
Durée : 3 ans
Contact : thao.dang@univ-grenoble-alpes.fr
Date limite de publication : 2022-08-31

Contexte :
The thesis is fully funded for three years by a grant from region Auvergne-Rhône-Alpes starting in 2022.
The description of the topic can be found at https://www.decyphir.com/PhD_Position_DETAI22.html

Cyber Physical Systems (CPS) are systems mixing software and hardware (cyber) components in interaction with their (physical) environment. Typical examples includ autonomous cars, robots, medical devices. Mathematically, they are modelled with so-called hybrid systems, which are dynamical systems with multiple modes, which can be continuous or discrete in nature. Since the modelisation includes the physical/biological environment, the models can be of arbitrary complexity, from trivial (not all models need be complex to be useful) to untractable for nowadays computational resources due to the infiniteness of input and state spaces of these systems. Hence new methods and tools are always needed to manage and handle the type of heterogeneous computations and data generated by the analysis and design of hybrid systems.

Sujet :
In this thesis, we want to tackle this issue from the angle of coverage measures. Given a CPS problem and some data and/or models (e.g., a hybrid system) associated to it, the question is: what is the mathematical domain that can represent all possible data that can be observed, and can we measure how well the given data represent this domain? This question is of primordial theoretical and practical interest in many contexts. One popular contemporaneous instance is that of machine learning (ML). It is well-known that ML-based algorithms, which are more and more used for CPS design, are only as good as the data used to train them. However it is much less well understood how to formally define the “goodness” of the data at our disposal. Hence there is a need for meaningful measures that can be computed and used not only to quantify the quality of a set, but also to fix it by, e.g., shrinking or augmenting it to better represent a domain to learn.
The questions of coverage, sampling, data augmentation, ML, CPS, etc are not new and topics that have attracted a lot of interest recently. The originality of this thesis will be to tackle these problems from the perspective of hybrid systems and formal methods, which are two research directions in which Verimag and Decyphir are specialized into and internationally recognized for. The intrinsic hybrid nature of data and systems considered in machine learning for CPS is often overlooked and we believed there is a need to study it in a more systematic and explicit way. Formal methods makes it possible to derive more rigorous guarantees and the hope is also that through the use of specification languages such as, e.g., Signal Temporal Logics (STL), they can help in the development of “explainable” measures, i.e., measures that are directly related to precisely formulated requirements as opposed to some hard to interpret mean squared error quantity as is the most frequent practice.

Profil du candidat :
We are looking for candidates with a Master degree in computer science or control engineering interested in CPS, artificial intelligence and machine learning. The thesis is expected to feature a strong experimental and development component but opportunities to developping theoretical contributions will also be likely. As a consequence, candidates with both theoretical and practical inclinations are welcome to apply.

Formation et compétences requises :
Master degree in computer science or control engineering

Adresse d’emploi :
Verimag Laboratory, Université Grenoble Alpes (UGA),
700 avenue centrale
38400 Saint Martin D’Hères

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

Laboratoire/Entreprise : Orange Lab / UCA-Inria-CNRS MAASAI Team
Durée : 36 months
Contact : michel.riveill@univ-cotedazur.fr
Date limite de publication : 2022-03-14

Contexte :
TThe goal of the thesis is to perform research on decentralized and efficient federated AutoML learning for heterogeneous embedded devices.

The training of AI models for service delivery is today facing a conceptual transformation, by shifting the learning of models close to the data, embedded on users’ devices. These devices have limited resources and must remain fully operational during the learning phase. In addition, users today generate sensitive data and new collaborative algorithms for learning need to be developed and optimized for different embedded devices, ranging from smartphones to IoT.

Nowadays, to build an AI model it is necessary to collect data on a central server (cloud). The problems of this method are related to privacy, control of data usage and computational resources. Federated learning (FL) [1,2] is a new AI approach with collaborative training that resolves these problems. Models are trained on local users’ data and its parameters only are exchanged with other users to build a global model. The challenges of Federated Learning are (a) obtaining efficient and robust decentralized FL models with heterogeneous data (b) optimizing resources for actual operational deployment and (c) customizing services and optimizing model based on available resources for groups of users, because a single global model may be less explainable, accurate and appropriate when compared to a personalized model.

We will deploy deep neural networks on users’ devices because they have high classification/prediction accuracy in various tasks. However, their training requires a significant effort in terms of finding optimal hyperparameters, which limits their use at devices with constrained resources. Emerging areas address the problem of automatic neural network generation [3] and automatic search for appropriate architectures (Neural Architecture Search-NAS), features required for real-world deployments. FL NAS [4] aims at optimizing the architecture of neural network models in the FL environment. Many questions in this domain remain open. For example, there are no approaches developed for FL with clients having the same sample space and a different feature space.

Sujet :
The objective of the thesis is to (a) design a federated learning framework to automatically generate low-power neural networks in compliance with GDPR [5] with homogeneous (b) and heterogeneous devices under device constraints (availability, resources, states) and to study it in a fully decentralized Peer-to-Peer federated learning setup.

[1] J. Konecny, H. B. McMahan, F. X. Yu, P. Richtarik, A. T. Suresh and D. Bacon, “Federated Learning: Strategies for Improving Communication Efficiency,” in arXiv, 2017, pp. 1-10.

[2] K. Bonawitz, H. Eichner, W. Grieskamp, D. Huba, A. Ingerman, V. Ivanov, Ch. M. Kiddon, J. Konečný, S. Mazzocchi, B. McMahan, T. Van Overveldt, D. Petrou, D. Ramage and J. Roselander, “Towards Federated Learning at Scale: System Design,” SysML 2019, https://arxiv.org/abs/1902.01046, 2019.

[3] A. Wong, M. J. Shafiee, B. Chwyl and F. Li, “FermiNets: Learning generative machines to generate efficient neural networks via generative synthesis,” 1809.05989.pdf (arxiv.org), NIPS, 2018.

[4] H. Zhu, H. Zhang and Y. Jin, “From Federated Learning to Federated Neural Architecture Search,” https://arxiv.org/pdf/2009.05868.pdf, 2020.

[5] Regulation (EU) 2016/679 of the European Parliament and of the Council (article 30), https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:32016R0679&from=EN#d1e3265-1-1, Archived from the original on 28 June 2017.

Profil du candidat :
You have a Masters degree in Data Science or Computer Science and you are a curious person that likes to learn and seek for solutions. You are highly motivated to do your thesis in the emerging field of distributed algorithms for embedded devices. You have skills in machine learning, optimization and statistics (essential) as well as good programming skills and knowledge in the field of embedded devices (desirable). Interest in the field of Signal Processing is a plus.

Furthermore, autonomy and open-mindedness are the qualities particularly appreciated for research work. The dynamism, the strength of proposal and the capacities of communication are also required for this position. English will be used throughout the thesis (reading state of the art, writing articles and presenting results at international conferences) and excellent level of English is therefore required.

Formation et compétences requises :
Master or Ecole d’ingénieur.

Adresse d’emploi :
Orange Lab Sophia Antipolis.

Contact :
– Tamara.TOSIC@orange.com,
– Michel.RIVEILL@univ-cotedazur.fr

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

Laboratoire/Entreprise : CRIStAL, Inria, Université de Lille
Durée : 3 ans
Contact : louis.bigo@univ-lille.fr
Date limite de publication : 2022-03-14

Contexte :
Depuis une dizaine d’années, les réseaux de neurones profonds font l’objet de nombreuses recherches dans le domaine du traitement automatique du langage naturel (Natural Language Processing). Ces recherches ont de multiples applications allant de l’analyse de corpus à la génération automatique de contenu.

La nature temporelle de la musique encourage et facilite sa représentation sous la forme de séquences d’éléments à différentes échelles, généralement des accords ou des notes, comparables à des séquences de mots. Cette séquentialité, ainsi que l’assimilation courante de la musique à une sorte de langage, ont motivé l’utilisation d’outils originalement conçus pour des tâches de NLP pour le traitement automatique de données musicales (Music Information Retrieval) pour des tâches variées incluant l’analyse et la génération automatique de musique.

Sujet :
L’objectif central de cette thèse est d’évaluer l’adaptabilité, la performance et la pertinence de techniques de NLP lorsqu’elles sont appliquées sur des données musicales. On se concentrera en particulier sur l’application en musique de trois principes essentiels du NLP :

* le principe d’attention mutuelle (self-attention)
* la segmentation (tokenization)
* l’apprentissage par transfert (transfert learning)

Ces principes seront étudiés à travers l’entraînement et l’évaluation de modèles musicaux inspirés par des modèles majeurs de NLP incluant l’auto-encodeur BERT ou le modèle auto-regressif GPT. Une réflexion sera menée sur les limites de l’application de modèles de langage naturel sur des données musicales, d’un point de vue technique comme d’un point de vue épistémologique, et sur les perspectives de modèles originaux spécifiquement adaptés à la modélisation de données musicales.

Plus de détails sur la page : http://www.algomus.fr/jobs/phd-nlp-en/

Profil du candidat :
Fort intérêt pour l’apprentissage automatique, le traitement automatique du langage naturel, et la musique.

Formation et compétences requises :
* Master d’informatique ou équivalent, apprentissage automatique, traitement automatique du langage naturel
* Connaissances et pratique musicales souhaitées.

Adresse d’emploi :
CRIStAL, Inria, Université de Lille, Villeneuve d’Ascq