Publications

Preprints and Accepted Papers

Preprints and papers from my PhD and earlier research. I focus on reducing computational cost without sacrificing model quality — whether that means smarter frame selection, sparse representations, or better understanding of robustness trade-offs.

Video understanding Knowledge distillation Adversarial robustness Efficient inference
2026
arXiv Preprint · in submission

PEEK: Picking Essential frames via Efficient Knowledge distillation

Killian Steunou, Anas Filali Razzouki, Mounîm A. El-Yacoubi, Khalil Guetari, Yannis Tevissen

We propose PEEK, a training-efficient frame selection method that distills knowledge from vision-language teachers to identify the most informative frames for video captioning. PEEK achieves competitive performance with significantly fewer processed frames, making dense video understanding more practical at scale.

Video Understanding Knowledge Distillation Frame Selection Video Captioning Efficiency
Project page arXiv PDF Code Model
2025
arXiv Preprint

Sparse Representations Improve Adversarial Robustness of Neural Network Classifiers

Killian Steunou, Théo Druilhe, Sigurd Saue

We revisit linear dimensionality reduction as a defense mechanism and show, both theoretically and empirically, that SPCA-based classifiers are more robust than PCA-based alternatives under adversarial attack. The analysis provides a principled explanation for when and why sparsity helps at the feature extraction stage.

Adversarial Robustness Sparse Representations Dimensionality Reduction Computer Vision
arXiv PDF Code

Coursework & Reproductions

Reports from my Master's — reproductions, adaptations, and applied projects.

2025 / Reproduction

Score-Based Generative Neural Networks for Large-Scale Optimal Transport

Original paper by Max Daniels, Tyler Maunu, Paul Hand

I reproduced the hybrid score-based transport approach and studied how regularization and sampling affect the quality of recovered maps.

Original paper Report GitHub
2025 / Adaptation

Test Time Training with Masked Autoencoders

Original paper by Yossi Gandelsman, Yu Sun, Xinlei Chen, Alexei A. Efros

I evaluated TTT-MAE on corruption benchmarks and explored an online variant that retains encoder updates between samples.

Original paper Report GitHub
2024 / Robustness

Are Generative Classifiers More Robust to Adversarial Attacks?

Original paper by Yingzhen Li, John Bradshaw, Yash Sharma

I revisited the robustness claims around generative classifiers and extended the original setup beyond MNIST to a more realistic dataset.

Original paper Report GitHub
2025 / Applied ML

Toxic Gas Characterization

Independent applied project

I studied domain shift caused by humidity changes and combined multi-task learning with adversarial adaptation for more stable gas characterization.

Report GitHub
2025 / Optimization

Convergence of SGD for Training Neural Networks with Sliced Wasserstein Losses

Original paper by Eloi Tanguy

We verified convergence behavior for sliced Wasserstein training on toy distributions and Fashion-MNIST, including a look at Noise Projected SGD.

Original paper Report GitHub
2025 / 3D Vision

An End-to-End Transformer Model for 3D Object Detection

Original paper by Ishan Misra, Rohit Girdhar, Armand Joulin

I reproduced 3DETR on SUN RGB-D and explored how an RGB-enhanced variant compares to the original point-cloud-only pipeline.

Original paper Report