Gradient-Free Warm-Start Library Recovery: an Amortized-Regret Separation

Jianwei Lou

Continual learning that is gradient-free, local, online, and append-only is attractive for edge and streaming deployment, but its value is usually argued informally. We give a prov…