Habib Slim

Hello! I am a freshly graduated C.S. engineer from ENSIMAG, a french Grande École specialized in applied mathematics and computer science.
I also earned a Master of Research in Data Science from Université Grenoble Alpes (UGA), during which I worked with Dr. Adrian Popescu on class-incremental learning for image classification at Université Paris-Saclay (CEA-LIST). You can read my master's thesis right here!
Prior to my M2, I worked with Pr. Georges Quénot on improving generative models for continual learning.

Email  /  CV  /  Master's thesis  /  Github  /  Twitter

Profile picture
Research

My research interests gravitate around computer vision, representation learning, transfer learning, adversarial ML and paradigms like continual/few-shot/zero-shot learning which all aim at making artificial neural networks closer to the learning abilities of biological systems.

Dataset Knowledge Transfer for Class-Incremental Learning without Memory
Habib Slim*, Eden Belouadah*, Adrian Popescu, Darian Onchis
WACV, 2022
pdf / code / bib

We look at ways to improve bias correction methods in continually learned deep classifiers, and enable their use when no rehearsal memory is available. We propose a finer modeling of prediction bias in incrementally learned models, and show that learned correction parameters are stable enough across data distributions to be transferable.
The resulting method largely improves the average accuracy of the state-of-the-art backbone methods tested.

Knowledge Transfer for Class-Incremental Learning without Memory
Habib Slim
Master's thesis, UGA - Grenoble INP, 2021
pdf / bib

My master's thesis on class-incremental learning in memoryless settings, conducted at Université Paris-Saclay (CEA-LIST). I obtained my Master of Research with the highest distinction (mention "très bien"), in front of a jury composed of Pr. Massih-Reza Amini, Pr. Franck Iutzeler and Pr. Ioannis Kanellos.

Boosting Generative Networks for Incremental Learning
Research Project, 2020
pdf / code

We study the potential of cGANs architectures, like Andrew Brock's BigGAN, as a pseudo-rehearsal method for incremental classifier learning. We show that when working on small-scale datasets, using data transformations on synthetic images enables the replacement of real data with GANs. Investigating the effects of the quality/diversity trade-off of generative models on the accuracy of trained classifiers, we show that quality is less important than diversity when training on synthetic data.

WIB: an integrated Wikipedia browser for participatory evaluation of relevance models
Christophe Brouard, Jean-Pierre Chevallet, Téo Orthlieb, Habib Slim
EGC, 2019
pdf (FR) / bib

A fun software engineering project I worked on at the end of my bachelor's degree. The project involved making a syntactic parser for Wikitext, the markup language of Wikipedia, which I wrote from scratch in ~10k lines of Java code. I also wrote a complete GUI to select articles to extract based on the Wikipedia ontology graph, and contributed to a web interface allowing to easily browse through Wikipedia articles. Our work resulted in a publication in the french EGC conference.

Fun stuff

I've selected below some other fun projects I've worked on recently.

Transcription factor binding prediction with kernel methods
Data Challenge, 2021
pdf / code

For this data challenge, our task was to predict whether input DNA sequence regions were binding or not to specific transcription factors We implement from scratch and compare various string kernels operating on DNA sequences, alongside SVM/KRR and KLR classifiers. I worked on this project with David Emukpere (INRIA) in the context of a course taught by Julien Mairal at UGA.

Distributed DNN training with OpenMPI
Software Project, 2020
pdf / code

This repository contains a minimal framework to quickly prototype deep architectures and facilitate weight and gradients sharing among processing nodes, written from scratch. We implement various distributed DNN optimization algorithms and conduct a benchmark of the different methods, evaluated on the MNIST and Fashion-MNIST datasets.

Fast decoding of JPEG-JFIF images
Software Project, 2019
code

A JFIF-compliant implementation of a JPEG image decoder, written entirely in C without any dependency and following the official ISO specification. This decoder supports JFIF files encoded in the sequential or progressive format, and implements the Loeffler algoritm for fast inverse DCT computation.

RoboKode: A serious game to learn programmation
Software Project, 2018
code (FR)

RoboKode is a 2D isometric serious game dedicated to learning programming and intended for high-school students. We introduced a simple programming language (with instructions in french) for an easy start, and guide the players through didactic puzzles slowly increasing in complexity. The game was programmed in Java and libGDX, a fully open-source 2D game development framework.
The game won the Professional Jury Award in a serious game competition at UGA in 2018.

Other stuff

I was the recipient of an Excellence Master Scholarship issued from the Agence Nationale de la Recherche (ANR).
I was a student tutor for a year (contract with UGA), during which I taught entry-level computer science courses (algorithmics and programming, computer architecture, arithmetics).
I also taught high-school level mathematics during two years on a volunteer basis, to students coming mostly from disadvantaged backgrounds.

Besides academia, I enjoy hiking and scuba diving (I'm a level 2 FFESSM diver). If you want to contact me, please feel free to shoot me an e-mail!


Credits to Jon Barron for the cool template! Last updated on October 2021.