Category: Non classé

The paper Exploiting Precision and Recall to assess the quality and diversity of LLMs, by Florian Le Bronnec, Alexandre Vérine, Benjamin Negrevergne, Yann Chevaleyre and Alexandre Allauzen has been accepted to the 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024)!

This paper adapts the precision and recall metrics used in image generation to the context of text generation, with experiments on large language models (LLMs) such as LLAMA-2 and Mistral. Congrats to all, and especially to Florian Le Bronnec and Alexandre Vérine for leading the effort!

Théo Delemazure, PhD student and MILES member, is first author on two papers accepted at the 33rd International Joint Conference on Artificial Intelligence (IJCAI 2024) :

– Selecting the Most Conflicting Pair of Candidates

– Comparing Ways of Obtaining Candidate Orderings from Approval Ballots

One paper authored by MILES member will be presented at AISTATS 2024, to be held on May 2-4 in Valencia, Spain:

Optimal Budgeted Rejection Sampling for Generative Models by Alexandre Vérine, Muni Sreenivas Pydi, Benjamin Negrevergne and Yann Chevaleyre.

Congratulations to Florian Le Bronnec, PhD student and MILES member, for receiving the best paper award at the 18th conference of the European chapter of the Association for Computational Linguistics (EACL 2024)!

This paper is a joint work with Alexandre Allauzen, as well as colleagues from Sorbonne Université, Criteo and Singapore.

Check out their work here: LOCOST: State-Space Models for Long Document Abstractive Summarization

MILES members are attending the Neural Information Processing Systems (NeurIPS) conference in New Orleans, LA, USA! Be sure to check their posters:

• On the Role of Randomization in Adversarially Robust Classification  by Lucas Gnecco Heredia, Muni Sreenivas Pydi, Laurent Meunier (former MILES member), Benjamin Negrevergne and Yann Chevaleyre
• Precision-Recall Divergence for Generative Modeling with GANs and Normalizing Flows by Alexandre Vérine, Benjamin Negrevergne, Muni Sreenivas Pydi and Yann Chevaleyre

Two papers co-authored by MILES members have been accepted to the 2023 edition of ICLR (International Conference on Learning Representations):

• Contextual bandits with concave rewards, and an application to fair ranking, by Virginie Do, Elvis Dohmatob, Matteo Pirotta, Alessandro Lazaric and Nicolas Usunier;
• A unified algebraic perspective on Lipschitz neural networks, by Alexandre Araujo (MILES alumni), Aaron J. Havens, Blaise Delattre, Alexandre Allauzen and Bin Hu.

Congratulations to all!

Six papers featuring members of MILES (and very recent alumni) were accepted at NeurIPS 2022, one of the flagship conferences in machine learning! Below are the papers (with MILES authors in bold)

•  Smooth Fictitious Play in Stochastic Games with Perturbed Payoffs and Unknown Transitions by Lucas Baudin and Rida Laraki
• An α-No-Regret Algorithm For Graphical Bilinear Bandits by Geovani Rizk, Igor Colin, Albert Thomas, Rida Laraki and Yann Chevaleyre
• Towards Consistency in Adversarial Classification by Laurent Meunier (recent MILES alumni), Raphael Ettedgui, Rafael Pinot (former MILES member), Yann Chevaleyre and Jamal Atif
• Benchopt: Reproducible, efficient and collaborative optimization benchmarks by Thomas Moreau et al, featuring Pierre Ablin (recent MILES alumni)
• Do Residual Neural Networks discretize Neural Ordinary Differential Equations? by Michael E. Sander, Pierre Ablin (recent MILES alumni) and Gabriel Peyré
• A framework for bilevel optimization that enables stochastic and global variance reduction algorithms by Matthieu Dagréou, Pierre Ablin (recent MILES alumni), Samuel Vaiter, Thomas Moreau.

Congratulations to all, and kudos to all authors who submitted at NeurIPS this year!

MILES has the great pleasure of hosting Dr. Makoto Yamada on September 22 for a guest seminar!

Title: Approximating 1-Wasserstein Distance with Trees

Abstract: Wasserstein distance, which measures the discrepancy between distributions, shows efficacy in various types of natural language processing (NLP) and computer vision (CV) applications. One of the challenges in estimating Wasserstein distance is that it is computationally expensive and does not scale well for many distribution comparison tasks. In this talk, I propose a learning-based approach to approximate the 1-Wasserstein distance with trees. Then, I demonstrate that the proposed approach can accurately approximate the original 1-Wasserstein distance for NLP tasks. (https://arxiv.org/abs/2206.12116)

Speaker bio: Makoto Yamada received the Ph.D. degree in statistical science from The Graduate University for Advanced Studies (SOKENDAI, The Institute of Statistical Mathematics), Tokyo, in 2010. Currently, he is a team leader at RIKEN AIP, an associate professor at Kyoto University, and a transitional associate professor at Okinawa Institute of Science and Technology (OIST). His research interests include machine learning and its application to biology, natural language processing, and computer vision. He published more than 50 research papers in premium conferences and journals such as NeurIPS, AISTATS, ICML, AAAI, IJCAI, and TPAMI, and won the WSDM 2016 Best Paper Award.

The seminar will take place on September 22 at 3pm in room A707 of Université Paris Dauphine-PSL.

MILES is thrilled to welcome Pierre Ablin as a new CNRS researcher within the team!

Pierre’s research interests revolve around optimization, machine learning, neural networks and brain signal processing. Stay tuned for more updates about Pierre’s research!

As part of its weekly group meeting, MILES is glad to announce three seminars in the upcoming weeks!

• September 30, 2021: In-person seminar by Ruben Ohana (ENS & LightOn) 
  Training neural networks with Direct Feedback Alignment: theory and applications in adversarial robustness
• October 7, 2021: In-person seminar by Laurent Daudet (LightOn)
  Optical random projections: scaling up randomized algorithms for scientific computing and machine learning
• October 14, 2021: Online seminar by Lindon Roberts  (ANU, Australia)
  Inexact Derivative-Free Optimization for Bilevel Learning

Want to give a seminar for our group and beyond? Please send an email at clement.royer@dauphine.psl.eu.