I'm a PhD student at IRIT working on provably robust deep learning for safety-critical autonomous systems (in partnership with SNCF). My research focuses on building AI systems with formal safety guarantees through mathematical constraints, efficient optimization, and uncertainty quantification.
- 🛡️ Adversarial Robustness — Certified defenses and Lipschitz-bounded networks
- ✅ Conformal Prediction — Reliable uncertainty quantification with formal guarantees
- ⚡ Optimization — Scalable methods for constrained networks and efficient training
- 🔐 Privacy — Differential privacy and private training methods
Triton-accelerated Newton-Schulz optimization (Turbo-Muon). Standalone implementations for PyTorch and Optax.
Fast differentially private training using Lipschitz networks. Eliminates clipping overhead in DP-SGD.
Scalable robust conformal prediction with finite-sample safety guarantees under adversarial perturbations.
Certifiably robust semantic segmentation methods.
⚙️ jaxlip — Under Development
Fast, compileable Lipschitz-constrained networks in JAX with efficient multi-GPU support.
📫 Reach me: thomasmassena@gmail.com | Website | Google Scholar