Rony Abecidan

A French engineering student very curious, perseverant and determined.

I am interested in computer vision, project managements, and video-editing.

Email

LinkedIn

GitHub

💻 Technical Experience

PhD Candidate in Data Science and Steganalysis @SIGMA (November 2021 - Today)

I am studying deep domain generalization to cope with heterogeneity in steganalysis

AI Research Scientist @IBM (June 2021 - October 2021)

MISSION : Evaluate the impact of unseen data on causal estimations for healthcare purposes

Data Scientist Intern @IQera (June 2020 - August 2020)

MISSION : Designing of an algorithm able to spot a relevant segmentation of debts portfolios.

PHASE 1 : Creation of a clean database from the raw data of the company

• Operations on databases using SAS and SQL
• Cleaning (Detection and handling of outliers, missing values and anomalies)
• Feature selection (Statistical tests : Chi-2 & Kruskal-Wallis, Random forest for estimating the influence of the features on the final decision).

PHASE 2 : Implementation of a clustering model in Python

• Unsupervised clustering (K-modes and K-prototypes algorithms ajusted to fit the needs of the company using an appropriate cost function)
• Use of decision trees for explaining the segmentations using rules on the features.
• Search of the best segmentation with successives iterations using random initializations.

💡 Research Project

🔍 Forgery Detection with domain adaptation @SIGMA (October 2021 - March 2021)

• Study of domain adaptative strategy in the context of forgery detection :

Through this project I worked on strategies enabling to generalize the knowledge that can obtain an intelligent algorithm on a distribution different but related to the training distribution.

This project was supervised by Jérémie Boulanger and Vincent Itier from the research team SIGMA, member of the CRIStAL research center at Lille

Librairies used : Numpy, Matplotlib, Pytorch-Lightning

More information on the project are available in my report

🎵 What are good convolutions for music analysis ? @Algomus (October 2019 - October 2020)

• Study of the impact of customized convolutional kernels for solving the key detection task.

Using different spaces for representing musical scores and, playing with customized kernels helped us to measure to what extent results obtained using deep learning strategies are consistent with the musical theory.

This project was supervised by Gianluca Micchi and Mathieu Giraud from the AlgoMUS research team at Lille

Librairies used : Numpy, Matplotlib, music21, Pytorch

More information on the project are available in my report

🖋️ Publications

• About Unsupervised Domain Adaptation for Image Classification (published on March, 2021)

Review paper written for the Spectra Competition launched by Mathpix in 2021

The bulk of machine learning models have a tendancy to rely too strongly to the distribution of the data on which they have been trained. Through this review paper I proposed to discuss about ways to design an image classifier able to generalize well on a different but related distribution from its training one.

• Towards Custom Dilated Convolutions on Pitch Spaces

Late-Breaking/Demo publication written (and accepted) for the conference ISMIR in 2020 - Cowritten with Mathieu Giraud and Gianluca Micchi

We benchmark several convolution kernels, in particular custom dilated convolutions. We test whether convolutions inspired by known pitch spaces like the Tonnetz may help to achieve better results on the key detection task.

🏆 Accomplishments

• Participation to a cartoon/short-movie festival at Lille, France in 2019 - Second regional award

• 11th grade Excellence Prize (STEM cursus), 2014

💬 Languages

French: Native
English: B2+
Spanish: B1+

🎓 Education

Msc. in Data Science - 2019/2021 - University of Lille, France

Some courses I liked

• Probability 2 (Recall of probability theory and study of Markov-chains)
• Statistics 1 & 2 (Descriptive and inferential statistics)
• Kernel Machines (RKHS, MMD)
• The landscape of machine learning (Theory behind classical ML algorithms)
• Deep Learning
• Graph-based Machine Learning (Clustering, Dimensionality Reduction and Semi-Supervised Learning with Graphs)
• Signal processing
• Bayesian Learning
• Distributed programming
• Sequential decision making/Reinforcement learning
• Privacy Preserving Machine Learning

This master is part of the « Information and Knowledge Society » Graduate Program which brings together 10 Master tracks (list), combining the expertises needed to build a human-friendly world. An outstanding scientific environment is provided with 13 research units recognized at the best international level.

Msc. in Multi-Disciplinary Engineering - 2017/2019 - Centrale Lille, France

Centrale Lille is one of the leading French graduate engineering schools.

Some courses I liked

• Mathematical analysis for engineers (Introduction to Lebesgue measure, Integration Theory, Distributions and Fourier Transform)
• Random modeling and scientific computing (Probabilities, Study about several kind of differential equations)
• Signals representation and Inverse problems (2D Fourier Transform, Fast Fourier Transform, Introduction to wavelets theory, Introduction to inverse problems)
• Optimization and prescriptive analysis (Simplex algorithm, linear programmation, integer linear programmation,primal&dual problems,etc…)
• Collaborative Intelligence (Genetic algorithm, Tabu, Simulated annealing, Conception of a multi-agents system able to resolve more efficiently the Travelling salesman problem than meta-heuristics)
• WEB 2.0 Technologies (Concepts for web developers, Introduction to HTML, CSS, Javascript,PHP, AJAX technologies, API Rest, mini-project which involved the use of NodeJS nd socket.io for the back-end part)

Bsc. in Mathematics - 2018/2019 - University of Lille, France

Degree obtained with highest honors