Chems Eddine Benaziza
GitHub (Contribution) [Merged on the main repo] [Final report]
TL;DR: CITRUS introduces a novel framework for learning from multidomain graph data using continuous PDE-based filtering. It addresses key limitations of discrete GNNs—such as over-smoothing and computational inefficiency—by modeling tensorial data on product graphs continuously. The framework is lightweight, stable, and scalable to any number of domains, unlocking new potential for real-world applications like traffic networks and multidimensional signals. Our contribution includes an application on another dataset and adding a pull request (that was merged to the main) to improve the open-source repository.
Chems eddine Benaziza Supervised by Pietro Gori
Kaggle Github [Paper]
TL;DR: In this project, we present an AlexNet-based architecture for classifying skin lesions, which can be deadly if cancerous. The pipeline involves preprocessing steps such as hair removal and physical data augmentation, followed by feeding the augmented images to a AlexNet architecture for classification. Without data augmentation, our model achieved a performance below 50%. However, with our proposed augmentation techniques, we significantly improved the accuracy to 0.69, resulting in a final ranking of 13/80 on a school Kaggle challenge
Chems eddine Benaziza, Sara Meziane, Mariia Baranova, Supervised by Arthur Leclaire
Github
TL;DR: By optimizing the orthogonal projection of our perturbation onto an existing perturbation in the feature space, we can improve transferability significantly. Choosing the layer at which we optimize the projection changes the transferability significantly. Pretty surprising that this works.This school project investigates the application of Maximum Mean Discrepancy (MMD) Generative Adversarial Networks (GANs) for image generation. We delve into the MMD loss formula, which measures the discrepancy between the distributions of real and generated images.
Our goal: is to compare MMD GANs with Wasserstein GAN (WGAN) to understand the trade-offs between various CNN feature extractors and the significance of gradient penalties.
Chems eddine Benaziza Supervised by Christophe kervazo
Github
TL;DR:
This work was done as a school project for the course Image Processing at Télécom Paris
IMA201.
The goal of this project is to generate superpixels using the watershed algorithm.
The implementation followed this
paper.
You will find the implementation in the Jupyter notebook, the Streamlit web app in app.py, and the
detailed report in
waterpixels_generation_report.