Worked in the Pytorch team, working on TorchRL and Tensordict.Focused on Model Based Reinforcement Learning models such as Dreamer.

Worked on Subject Transfer and Conditional Generative Adversarial NetworksInternship done at two labs:- IMAGINE, Research Group in Computer Vision, Machine Learning and Optimization, ENPC- GeoVic: Research Group in Computer Vision and Computer Graphics, École Polytechnique

Worked on Pose Transfer and Conditional Generative Adversarial Networks Internship done at two labs: - IMAGINE, Research Group in Computer Vision, Machine Learning and Optimization ENPC - GeoVic: Research Group in Computer Vision and Computer Graphics, École Polytechnique

Worked on detecting fraudulent patterns between traders with anomaly detection for one of the largest Asian banks. The model had to work in an unsupervised manner and handle high volume of data. Researched state-of-the-art techniques such as Graph Neural Networks (GraphSAGE) and its variations. Worked on an autoencoder-based embedding.Implemented models using Pytorch and the torch-geometric module for the graph embedding. Collaborated on an Upskills internal project. Developed a supervised model using logistic regression and a novelty detection system based on Variational Autoencoder (VAE).

Supervised by Clément FERNANDEZ and Wojciech PIECZYNSKI. Worked on an image segmentation algorithm based on a semi-Markov hidden model. This graphical model learn the latent waiting time in each class of the image removing the geometric assumption of the Markov Chain models. Adapted and implemented an SEM algorithm to estimate such parameters. Optimized computation time using Numpy.

Research project supervised by Sylvain LECORFFTheoretical study on the selection of better activation function and initialization of our neural network. Used Gaussian Processes to model an infinitely wide neural network (Also known as Neural Tangent) and study its properties. Gained exposure to cutting edge research papers and developed rigorous mathematical proofs. Vulgarized the technical aspect of this project to a non-expert audience.Based on the article: Soufiane Hayou, Arnaud Doucet, & Judith Rousseau. (2019). On the Impact of the Activation Function on Deep Neural Networks Training.

Worked on a better way to store genotypic data. Tools such as Hadoop, Spark and the Parquetfile format were used to optimize the storing performance. Developed some custom algorithms inPython to fit the data needs. Presented the results by a poster to a wide audience of researchers.