Yannis Flet-Berliac

Postdoc at Stanford

Stanford AI Lab

Hello, I am a Postdoctoral Scholar at Stanford Artificial Intelligence Laboratory where I work with Emma Brunskill. I am particularly interested in Offline Reinforcement Learning and Deep RL with applications in Healthcare and Robotics.

I did my PhD at Inria where I worked on sample-efficient Deep RL for control, exploration and safety. I graduated from DTU (machine learning, deep learning, HPC) and École Centrale (advanced probabilities, software engineering, project management). I also worked as a ML Engineer at iAdvize (NLP) and in three different startups (NLP, computer vision, speech processing) in Denmark.

Whenever possible, I like to go out running (at times marathons), cross-country skiing and windsurfing.

Interests

Artificial Intelligence & Society
Reinforcement Learning
Machine Learning
Deep Learning
Producing music & videos
Fly drones, play the piano, …

Education

PhD in Computer Science, 2021

Inria (funded by Univ. Lille), Lille, FR
MSc in Computer Science, 2017

Technical University of Denmark, Copenhagen, DK
MSc in General Engineering, 2017

École Centrale, Nantes, FR
Degree in Digital Business and IT, 2015

Audencia School of Management, Nantes, FR
BSc in General Engineering, 2014

École Centrale, Nantes, FR

Publications

More Publications

A. Nie, Y. Flet-Berliac, D. Richmond, W. Steenbergen, E. Brunskill
Data-Efficient Pipeline for Offline Reinforcement Learning with Limited Data.
36th Conference on Neural Information Processing Systems (NeurIPS), 2022.

Y. Liu, Y. Flet-Berliac, E. Brunskill
Offline Policy Optimization with Eligible Actions.
38th Conference on Uncertainty in Artificial Intelligence (UAI), 2022.

Link PDF

Y. Flet-Berliac, D. Basu
SAAC: Safe Reinforcement Learning as an Adversarial Game of Actor-Critics.
5th Conference on Reinforcement Learning and Decision Making (RLDM), 2022.

Link PDF PDF (extended version)

Y. Flet-Berliac
Jury: Ann Nowé, Bruno Scherrer, Luce Brotcorne, Anders Jonsson, Joëlle Pineau, Adam White
Sample-Efficient Deep Reinforcement Learning for Control, Exploration and Safety.
PhD Thesis, 2021.

Link PDF

Y. Flet-Berliac, J. Ferret, O. Pietquin, P. Preux, M. Geist
Adversarially Guided Actor-Critic.
9th International Conference on Learning Representations (ICLR), 2021.

Link PDF

Y. Flet-Berliac, R. Ouhamma, O.-A. Maillard, P. Preux
Learning Value Functions in Deep Policy Gradients using Residual Variance.
9th International Conference on Learning Representations (ICLR), 2021.

Link PDF

Y. Flet-Berliac, P. Preux
Only Relevant Information Matters: Filtering Out Noisy Samples to Boost RL.
29th International Joint Conference on Artificial Intelligence (IJCAI), 2020.

Link PDF

J. Demongeot, Y. Flet-Berliac, H. Seligmann
Temperature Decreases Spread Parameters of the New Covid-19 Case Dynamics.
Biology, 9(5), p.94, 2020.

Link PDF

Y. Flet-Berliac, P. Preux
MERL: Multi-Head Reinforcement Learning.
33rd Conference on Neural Information Processing Systems (NeurIPS), Deep RL Workshop, 2019.

Link PDF

Y. Flet-Berliac, P. Preux
High-Dimensional Control Using Generalized Auxiliary Tasks.
Research Report hal-02295705, 2019.

Link PDF

Invited Talks

Efficient Actor-Critics under the Prism of Variance

July, 2021

UC Berkeley Robot Learning Lab

Adversarially Guided Actor-Critic

February, 2021

DeepMind

Slides

Learning Value Functions using Residual Variance in Deep Policy Gradients

October, 2020

DeepMind

Slides

Oral & Panel Discussion: Do we control the algorithms we create?

November, 2019

Open Forum Art and Research

Improving Policy Gradient Updates with MERL and SAUNA

October, 2019

Inria Seminars (SequeL)

Deep Reinforcement Learning at Scale

April, 2019

HPC - BigData Inria Project Lab

QA and Deep Learning for Language Understanding

November, 2017

Machine Learning Meetup

Selected Software

rlberry

A Reinforcement Learning Library for Research and Education (PyTorch)

adversarially-guided-actor-critic

AGAC: Adversarially Guided Actor-Critic (TensorFlow)

actor-with-variance-estimated-critic

AVEC: Actor with Variance Estimated Critic (TensorFlow)

rlss-2019

Materials for the Reinforcement Learning Summer School 2019: Bandits, RL & Deep RL (PyTorch)

Teaching

Reinforcement Learning - Fall 2019 - MVA - ENS Paris-Saclay

Teaching Assistant
Instructors: Alessandro Lazaric, Matteo Pirotta

Reinforcement Learning Summer School 2019

Teaching Assistant
Instructors: Felix Berkenkamp, Tristan Cazenave, Ludovic Denoyer, Gabriel Dulac-Arnold, Audrey Durand, Vincent François-Lavet, Matteo Hessel, Emilie Kaufmann, Marc Lanctot, Max Lapan, Alessandro Lazaric, Odalric-Ambrym Maillard, Jérémie Mary, Gerhard Neumann, Guillaume Obozinski, Olivier Pietquin, Bilal Piot, Matteo Pirotta, Bruno Scherrer, Florian Strub, Eleni Vasilaki, Oriol Vinyals

Professional Experience

January 2022 – Present

Palo Alto, CA, USA

Postdoctoral Scholar

Stanford University

November 2017 – October 2018

Nantes, FR

Machine Learning Engineer

iAdvize

Designed and executed product-focused research agendas, which led to building a conversational model for human/machine interface using deep learning.

August 2017 – November 2017

Copenhagen, DK

Research Assistant

DTU

Research work at DTU Compute laboratory focusing on deep convolutional neural network models for image classification and generative adversarial network models for image generation from a mixture of human artworks and photographs.

March 2017 – August 2017

Copenhagen, DK

Machine Learning Researcher

Soply (part-time during MSc)

Defined with the co-founders a roadmap for ML projects in the company, which led to building a system to recommend artists according to their photographic style and three artworks classification models (content, style & type) in collaboration with the National Gallery of Denmark.

November 2015 – January 2017

Copenhagen, DK

Machine Learning Engineer

EasyTranslate (part-time during MSc)

Several research projects in collaboration with the product team including a seq2seq machine translation model for specialized text and a recommendation system for human translators using LDA models trained on Wikipedia, deployed on AWS.