Andrei Cozma’s research asks whether the tools of dynamical-systems theory can make reinforcement learning less fragile. His master’s thesis — completed at UT in 2024 — became KIPPO: Koopman-Inspired Proximal Policy Optimization, published at IJCAI 2025 with Landon Harris and Dr. Qi, which uses Koopman-operator structure to smooth and stabilize policy optimization.
He is also a co-author of Cross-Scale MAE (NeurIPS 2023), the lab’s multi-scale self-supervised pre-training framework for remote sensing, and has worked on industrial computer vision including defect detection in tire X-ray imagery. He continues in the PhD program, working across the lab’s learning-dynamics and representation-learning threads.
Projects
Koopman Dynamics
Treating neural network training as a dynamical system — predicting weights epochs ahead to accelerate optimization, and stabilizing policy learning in RL.
Representation Learning
From Cross-Scale MAE to ExPLoRe — a sustained line of work on learning visual representations without labels, published at NeurIPS 2023 and ECCV 2026.
Publications
KIPPO: Koopman-Inspired Proximal Policy Optimization
International Joint Conference on Artificial Intelligence (IJCAI) 2025
Defect Detection in Tire X-Ray Images: Conventional Methods Meet Deep Structures
arXiv 2024
Cross-Scale MAE: A Tale of Multiscale Exploitation in Remote Sensing
Advances in Neural Information Processing Systems (NeurIPS) 2023