Double Preconditioning (DoPr): Optimization for Test-Time Performance, not Validation Loss

arXiv:2606.06418v1 Announce Type: cross Abstract: Many modern applications of deep learning involve training a neural network via a one-step prediction loss (e.g., $L^2$ regression, cross-entropy), but deploy the network by rolling out along its own predictions. Key examples include autoregressive language modeling, flow-based generative...

Läs hela artikeln hos källan.