2 papers
cs.LG2026
Pro-KLShampoo: Projected KL-Shampoo with Whitening Recovered by Orthogonalization
Ruotong Sun, Ermin Wei
Optimizers that exploit the matrix structure of gradients are central to modern LLM pre-training, with two distinct frontiers: explicit Kronecker-factored preconditioning -- most r…
cs.LG2025
Hessian-Free Online Certified Unlearning
Xinbao Qiao, Meng Zhang, Ming Tang +1
Machine unlearning strives to uphold the data owners' right to be forgotten by enabling models to selectively forget specific data. Recent advances suggest pre-computing and storin…