Open Research Positions

The ANR project E-SSL (Efficient Self-Supervised Learning for Inclusive and Innovative Speech Technologies) has started this year. Self-supervised learning (SSL) has recently emerged as one of the most promising artificial intelligence (AI) methods as it becomes now feasible to take advantage of the colossal amounts of existing unlabeled data to significantly improve the performances of various speech processing tasks.

Speech technologies are widely used in our daily life and are expanding the scope of our action, with decision-making systems, including in critical areas such as health or legal aspects. In these societal applications, the question of the use of these tools raises the issue of the possible discrimination of people according to criteria for which society requires equal treatment, such as gender, origin, religion or disability… Recently, the machine learning community has been confronted with the need to work on the possible biases of algorithms, and many works have shown that the search for the best performance is not the only goal to pursue [1]. For instance, recent evaluations of ASR systems have shown that performances can vary according to the gender but these variations depend both on data used for learning and on models [2]. Therefore such systems are increasingly scrutinized for being biased while trustworthy speech technologies definitely represents a crucial expectation. Both the question of bias and the concept of fairness have now become important aspects of AI, and we now have to find the right threshold between accuracy and the measure of fairness. Unfortunately, these notions of fairness and bias are challenging to define and their meanings can greatly differ [3].

The goals of this PhD position are threefold:

• First make a survey on the many definitions of robustness, fairness and bias with the aim of coming up with definitions and metrics fit for speech SSL models
• Then gather speech datasets with high amount of well-described metadata
• Setup an evaluation protocol for SSL models and analyzing the results.

• Master 2 in Natural Language Processing, Speech Processing, computer science or data science
• Good mastering of Python programming and deep learning framework.
• Previous experience in bias, fairness and privacy in machine learning would be a plus
• Very good communication skills in English
• Good command of French would be a plus but is not mandatory

The PhD position will be co-supervised by Alexandre Allauzen (Dauphine Université PSL, Paris) and Solange Rossato and François Portet (Université Grenoble Alpes). Joint meetings are planned on a regular basis and the student is expected to spend time in both places. Moreover, two other PhD positions are open in this project. The students, along with the partners will closely collaborate. For instance, specific SSL models along with evaluation criteria will be developed by the other PhD students. Moreover, the PhD student will collaborate with several team members involved in the project in particular the two other PhD candidates who will be recruited and the partners from LIA, LIG and Dauphine Université PSL, Paris. The means to carry out the PhD will be provided both in terms of missions in France and abroad and in terms of equipment. The candidate will have access to the cluster of GPUs of both the LIG and Dauphine Université PSL. Furthermore, access to the National supercomputer Jean-Zay will enable to run large scale experiments.

Applications must contain: CV + letter/message of motivation + master notes + be ready to provide letter(s) of recommendation; and be addressed to Alexandre Allauzen (alexandre.allauzen@espci.psl.eu), Solange Rossato (Solange.Rossato@imag.fr) and François Portet (francois.Portet@imag.fr). We celebrate diversity and are committed to creating an inclusive environment for all employees.

The following topics are proposed for funded internship positions. Most of them can also be extended by a funded PhD position.

• On the Effectiveness of Explainability Methods for Neural Natural Language Processing
• Stable and 1-Lipshitz architectures for NLP
• Data2Text: learning to generate text from table
• Tranformers for speech processing

If you are interested in, feel free to contact me.

The interaction between machine learning and Physics has recently emerged as a new and important research area. Some illustrations are simulations of complex physical systems with machine learning models, or at the opposite, the introduction of numerical methods in machine learning.

At the interfaces of artificial and Physics, different tracks described below can be explored depending on the skills of the candidate.

Within the LAMSADE, we have started to work on the study of stability and robustness of neural network architectures. Deep learning algorithms have shown their vulnerability to adversarial attacks: small and imperceptible perturbations of the inputs that maliciously fools the prediction. Since this discovery, building adversarial attacks and defenses against them have been an active and hot research topic. However, the understanding of this phenomenon remains in the early stages and their is still a gap to come up with to better understand adversarial attacks.

Recent papers have explored the interaction between machine learning and numerical methods, opening new perspectives for many questions. For instance, this paper proposes an interpretation of deep learning as a parameter estimation problem of nonlinear dynamical systems. This formulation allows us to analyze the stability and well-posedness of deep neural network architectures like ResNet.

The goal of this 3 months is to develop a new kind of architecture, based on either the recurrent models or the transformers. The goal is to design a more robust and stable architecture. The development relies on pytorch and the experiments will use standard datasets (i.e NLP or image classification) and datasets from Physics.

Requirements:

• Outstanding master's degree (or an equivalent university degree) in computer science or another related disciplines (as e.g. mathematics, information sciences, computer engineering, etc.).
• Proficiency in machine learning, computer vision, or signal processing.
• Fluency in spoken and written English is required.
• Strong background in deep-learning with pytorch.

Application: To apply, please email alexandre.allauzen [at] dauphine.psl.eu with:

• a curriculum vitae
• a cover letter
• a research outcome (e.g. master thesis and/or published papers) of the candidate
• a transcript of grades

The interaction between machine learning and Physics has recently emerged as a new and important research area. Some illustrations are simulations of complex physical systems with machine learning models, or at the opposite, the introduction of numerical methods in machine learning.

At the interfaces of artificial and Physics, different tracks described below can be explored depending on the skills of the candidate.

In modern machine learning, the cornerstone is to let the model learn its own representation of the process from data observation. While, for many applications, data are readily available (computer vision, natural language processing, . . . ), some requirements are not met in the case of complex physical systems. Without loss of generality, let us consider the example of a turbulent flow field or the prediction of the sea surface temperature. The corresponding dataset is really small and scarce compared with usual machine learning applications. More importantly, the state cannot be fully observed in many situations and the data acquisition step often introduces noise.

The issues raised by noisy and scarce dataset are not new in the machine learning domain and there is, for instance, a long history of research in the field of generative models and how to represent high dimensional datasets in a compressed mathematical model. However, in the context of Physics, we can leverage some important properties like symmetries and invariances to address these challenges.

In some cases, a mathematical model for the system at hand is available, for instance: dynamical systems such as the Lorentz (63 and 93) attractors, Kuramoto-Sivashinsky and Kardar-Parisi-Zhang. With these case studies, this step includes the important definitions of the physical properties we want to introduce in the machine learning models.

Two approaches can be considered:

• Physical regularization: the loss function optimized during the training process can be augmented with tailored regularization terms. As an example, optimal transport-based (OT) loss definitions are often more relevant for physical systems featuring significant structure. This will be made computationaly tractable with the convolutional Wasserstein flavor of OT, e.g., see this paper.
• Adversarial training: the second approach relies on the recent adversarial learning trend to guide the model during the training process toward solutions that exhibit the desired properties. Early efforts are reported in this paper where the solution and the test functions in the weak formulation of high-dimensional linear and nonlinear PDE problems are parameterized as a primal and adversarial networks respectively.

The relationship between neural networks and differential equations has been studied in several recent works Lu et al. (2018); Chen et al. (2018). In particular, the very efficient neural architecture ResNet (or Residual Network) can be interpreted as discretized ordinary differential equations. This kind of architectures leads to a very large number of parameters. Hence, while the idea is really appealing in our context, architectures like ResNet are suitable for applications beyond our scope, where the data availability is not an issue. Pushing the discretization step towards its limit of zero, along with parameters tying, have given rise to a new family of models called Neural Ordinary Differential Equations (or Neural ODEs). In these recent papers and their extension, Dupont et al. (2019), the experimental setup mainly relies on conventional datasets used in image classification (MNIST or CIFAR10). Preliminary work on this new type of neural networks has demonstrated its parameter efficiency for supervised learning task which can be of a great importance in our case.

Applications can be sent electronically and should include a cover letter, full CV, and eventually references. A first round of interviews will start in first week of June 2021 so please submit as soon as possible. Feel free to contact us if you have questions on the topic and the position.

Alexandre Allauzen: alexandre.allauzen@dauphine.psl.eu Sergio Chibbaro: sergio.chibbaro@sorbonne-universite.fr

"When deep learning meets ergodic theory", M.A Bucci, O.S emeraro, S. Chibbaro, A. Allauzen, L. Mathelin, in https://hal.archives-ouvertes.fr/LIMSI/hal-03101431v1

"Control of chaotic systems by deep reinforcement learning", M.A Bucci, O.S emeraro, S. Chibbaro, A. Allauzen, et al. https://hal.archives-ouvertes.fr/LIMSI/hal-02406677v1

"Hamiltonian Neural Networks", Samuel Greydanus, Misko Dzamba, Jason Yosinski, NeurIPS 2019 proceedings. https://papers.nips.cc/paper/2019/hash/26cd8ecadce0d4efd6cc8a8725cbd1f8-Abstract.html