|
In-Context Learning as causal structure learners
Dec' 24
Work in progress...
Keywords : Causal Structure Learning, Compression, Meta-learning
|
 |
In-Context Learning and Occam's razor
Mar' 24 - Oct' 24
We propose an explanation for the strong generalization abilities of
in-context learners at inference time, by drawing connections between
meta-learning, In-Context Learning and Information theory.
Keywords : Meta-learning, Kolmogorov complexity, Occam's razor.
Paper
(Under review) /
Code
|
 |
The BrowserGym Ecosystem for Web Agent Research
BrowserGym is a fully deployed Open-Source Gym environment for Web-Automation. It serves to evaluate Web
Agents at solving a wide range of tasks on the web. it's a common effort between Service Now Research, CMU
and Mila.
Keywords : Web-Automation, Task solving, benchmark.
Journal Paper
(Accepted at TMLR) /
Code
|
 |
WorkArena : LLMs as Generalist Web Agents
Avr' 23 - Mar' 24
As part of my research at ServiceNow Research, I worked on WorkArena : an
Open-Source benchmark as a Gym environment to evaluate Web Agent to solve common knowledge
task on a Web Browser. Published at NeurIPS 2023 FMDM Workshop and ICML 2024.
Keywords : Web-Automation, Task solving, benchmark.
Paper
(Accepted
at ICML2024) /
Website
/
Code
Workshop Paper
(Accepted
at FMDM@NeurIPS2023)
|
 |
SeaPearl : a Constraint Programming Solver guided by
Reinforcement Learning
Feb' 21 - Jul' 24
I was an active lead maintainer of the Open-Source and collaborative
project SeaPearl. The goal is to develop an intelligent constraint
programming solver that can learn to solve any Constraint Optimization
Problems using Reinforcement Learning on Graphs.
The entire solver is written in Julia  . Click
here for detailed
explanations.
Keywords : Combinatorial Optimization, Reinforcement learning, GNNs
Journal Paper
(Accepted
at Constraint)
Original Paper (Accepted
at CP2023)
/
Code
|
 |
Adversarial Attacks on Sentiment Analysis models
Mar' 22 - May' 22
This project was carried out in the framework of the IFT6167 seminar led by Prof. Irina
Rish (Mila, Québec).
In this work, we aim to show, regarding Natural Language
Sentiment Analysis, that there exists a relationship between model size and
robustness to adversarial attacks. Ultimately, uncovering the emergence of power laws and
testing the robustness of language model with scale. We evaluate the performance of various
Eleuther AI GPT
models such as GPT-Neo 125M, GPT-Neo 1.3B, GPT-Neo 2.7B, GPT-J 6B against adversarial
attacks. We fined-tuned (trained on adversarial example) our different GPT models on common
datasets (Rotten Tomatoes, IMDB...) and evaluated them separately to quantify the effects of
scale on adversarial training.
Report /
Slides
|
 |
Diffusion Geodesic : a new Metric for non-linear Dimensionality Reduction
Sep' 21 - Jan' 22
In collaboration with Ph.D. candidate Guillaume Huguet (Mila,
Québec), we present our method for non-linear dimensionality reduction called Diffusion Geodesic.
Dimensionality reduction techniques are often used to visualize the underlying geometry of a
high-dimensional dataset. These methods usually rely on specific similarity measures. In this project,
we first approximate the geodesic distance using a diffusion process over the underlying manifold, then
we use Multi-Dimentionnal Scaling combined with our previously defined pairwise
'distances' to embed our Manifold in a lower dimensional space. We compare our model with popular
algorithms such as PHATE, UMAP, and Isomap on toy datasets and RNA-seq dataset.
Report /
Code
|
 |
Autonomous Drone Swarm Deployment
Dec' 20 – Mar' 21
In collaboration with Sariah Al Saati (ENS), Mehdi Benharrats (X-HEC), Swann Chelly (Sorbonne
University) and Pierre Tessier, this report proposes a method for the coverage of a rescue zone with a
swarm of UAV’s in order to detect possible target of interest.
The method is based on Collaborative Reinforcement Learning. It also presents a pipeline to locate
points of interest in 3D from a set of 2D images using Inverse Projection Transformation and 3D ray clustering.
Report
|
 |
Realtime 3D Deep Motion Capture
Oct' 20 – Dec' 20
In collaboration with Pierre Tessier (MS, Columbia University), the objective of our project was to
implement a intelligent 2D to 3D Motion Capture mechanism that uses only the video stream of a webcam as
input. We were able to animate relatively accurately this Mk-44 Iron Man 3D model.
The project is based on the model DOPE presented in this paper for the automatic 3D rig
generation from video input coupled with a quaternion-based 3D rotation inference pipeline for 3D model animation.
Report (french) /
Video /
Code /
Slides
|
 |
Sketch-based Shape Retrieval
Sep' 20 – Dec' 20
In collaboration with Pierre Tessier (MS, Columbia University), the objective of our project is to
enable an efficient search in a 3D model bank models from simple hand drawings. The project is based on
this paper SIGGRAPH2012.
The project lies on differents techniques such Suggestive Contouring (Canny filter), Gabor filtering
(gaussian convolution on Fourier Transformation of the input image) and histogram representation using
Visual Vocabulary.
Looking back at this project, I saw that another paper with much more impressive results came out a few years later using
Siamese convolutional Neural Networks for feature extraction.
Report (french) /
Code /
Slides
|
 |
Procedural modeling of a 3D rendered scene
Mar' 20 – June' 20
In collaboration with Elsa Deville  (MS, Imperial College), the
objective of our project is to fully render a realistic 3D marine scene using only randomized procedural
modeling (OpenGL).
The project implements different visual elements among them terrain generation using Perlin Noise, a realistic simulation of a fish swarm movement based on this
paper and realistic-looking ocean waves. (Trochoidal Waves : exact
solution of the Euler equations for periodic surface gravity waves).
Report (french)
|
