As I approach the final year of my PhD, I am open to discussing new opportunities as an AI Researcher or AI Software Engineer.
Feel free to reach out at thomasmassena@gmail.com.
News
-
Our paper, From SGD to Muon: Adaptive Optimization via Schatten-p Norms, was accepted at NeurIPS 2026.
-
Our paper, Fast and Flexible Robustness Certificates for Semantic Segmentation, was accepted at ECCV 2026.
-
Merged a PR on the Optax repository containing coefficient presets and various normalization schemes for the Muon optimizer, along with the Turbo-Muon contributions.
Recorded Talks
I was invited to pitch my recent paper, “Fast and Flexible Robustness Certificates for Semantic Segmentation,” at ANITI Days 2026 in Toulouse. You can watch the ANITI Days talk and the associated five-minute ECCV presentation.
Awards
Together with my coworker Leo Andeol, I received the Alexey Chervonenkis Award for Best Poster at the Fourteenth Symposium on Conformal and Probabilistic Prediction with Applications (COPA 2025).
Publications
From SGD to Muon: Adaptive Optimization via Schatten-p Norms NeurIPS 2026
Initially developed in this blog post, we show that a first-order model of the LMO update rule's optimality is sufficient to match or improve upon the better-performing optimizer between Adam or Muon across diverse training tasks.
Paper →Turbo-Muon: Almost-Orthogonal Pre-Conditioning for Fast Muon Updates Pre-print, under review
We improve the efficiency of the costly Newton-Schulz iteration of the Muon optimizer by using a preconditioning method from the Approximately Orthogonal Layer paper from Prach et al. This allows us to conserve the impressive performance of the Muon optimizer while gaining substantial computational efficiency at scale.
Paper →Fast and Flexible Robustness Certificates for Semantic Segmentation ECCV 2026
We use Lipschitz neural networks to perform certifiably robust segmentation tasks on challenging datasets such as CityScapes. Our networks are approximately 600 to 2000 times more computationally efficient at inference time. We additionally develop a full framework for the certification of complex deep learning models under arbitrary threats under two different paradigms.
Paper →Efficient Robust Conformal Prediction via Lipschitz-Bounded Networks ICML 2025
We provide a method to enable ~1000x more memory efficient Robust Conformal Prediction compared to related works without any performance loss. This enables efficient prediction with guaranteed error rates in noisy or adversarial environments.
Paper →DP-SGD without Clipping: The Lipschitz Neural Network Way ICLR 2024
We show that Lipschitz-constrained neural networks allow fast and intuitive Differentially Private training, reducing DP-SGD training time significantly and eliminating detrimental clipping bias.
Also presented as an invited Google Tech Talk.
Paper →An Adaptive Orthogonal Convolution Scheme for Efficient and Flexible CNN Architectures ICML 2025
We implement fast and flexible parametrizations for orthogonally constrained convolutions that match the time and memory consumption of unconstrained convolutions in large batch size settings.
Paper →Sequential Conformal Risk Control for Safe Railway Signaling Detection COPA 2025
We enable safe railway signaling detection with risk control guarantees on detection confidence, localization, and classification.
Paper →