DiffusionBlocks: Training Neural Networks One Block at a Time (opens in new tab)
A principled framework that converts a residual network into independently trainable blocks via a diffusion interpretation, achieving B× memory reduction without sacrificing performance.
Read the original article