pub.sakana.ai
DiffusionBlocks: Training Neural Networks One Block at a Time
A principled framework that converts a residual network into independently trainable blocks via a diffusion interpretation, achieving B× memory reduction without sacrificing performance.