Hallison Paz

Research Scientist based in France working at the intersection of adversarial machine learning, robust learning, and visual intelligence. My research focuses on understanding and defending against vulnerabilities in machine learning systems, while my broader technical foundation spans representation learning, generative modeling, and computer vision, with experience taking research ideas into production-grade systems.

I'm currently a Postdoctoral Researcher at the Centre de Mathématiques Appliquées (CMAP), École Polytechnique, working on data poisoning and dataset ownership verification methods in El Mahdi El Mhamdi's group. Previously, at Hand Talk by Sorenson, I led applied research on multimodal foundation models for bidirectional sign language translation (Libras and ASL), combining vision, language, and temporal modeling for human-centered AI.

I hold a PhD from IMPA (Brazil) and have published at SIGGRAPH, CVPR, and Computers & Graphics. I previously worked as a Research Scientist Intern at Meta Reality Labs, contributing to photorealistic human avatars, neural rendering, and 3D generative modeling.

Research Interests

• Adversarial Robustness & Decision Geometry: Understanding how high-dimensional decision boundaries interact with data manifolds and developing certified defenses against perturbations.
• Neural Implicit Representations & Coordinate Networks: Signed distance functions (SDFs), neural radiance fields, continuous surface morphing, and level-set geometry.
• Open Science & AI Education: Building accessible, mathematically rigorous machine learning educational resources.

Science Communication & Outreach

Beyond academia, I am the co-founder of Programação Dinâmica, the largest YouTube channel in Brazil dedicated to AI and data science, with over 240,000 subscribers and millions of views. I create technical content on programming, machine learning, and research to help grow the field in Brazil and inspire the next generations.