Thomas Fel

Thomas Fel is a Research Fellow at the Kempner Institute, Harvard University, specializing in Explainable AI. His work integrates computational science, mathematics, and neuroscience principles. Previously, he completed a PhD at Brown University under Thomas Serre and contributed to the DEEL project at Toulouse University with SNCF support. He has also interned at Google and GoPro.

Research Interests focus on advancing model interpretability through interdisciplinary approaches. Key projects include developing novel sparse autoencoder architectures (MP-SAE, RA-SAE, USAE) and creating benchmarks like Visual Anagrams to evaluate holistic shape processing in vision models. His work addresses generative model limitations, feature visualization stagnation, and conceptual blindspots in AI systems.

Scientific Recognition:

• 2025 AFIA National PhD Thesis Award
• 2025 'Signal, Image, Vision' Best PhD Thesis Award
• 2025 ICML Top Reviewer

Thomas has developed critical tools like LENS Project for feature visualization and Xplique Toolbox for attribution methods. His recent work at ICML 2025 explores hierarchical concept extraction and cross-modal representation alignment.