Orthogonal Subspace Fine-Tuning: freeze 99.3% of pretrained GAN weights, fine-tune only the residual subspace. Discovered 99.5% of physics gradients attack the principal SVD subspace (η=0.005, 40× stronger than predicted).
deep-learning pytorch gan svd topology-optimization fine-tuning peft continual-learning physics-informed orthogonal-subspace
-
Updated
Jun 12, 2026 - Python