Sakana AI Presents DiffusionBlocks: A Method for Training Neural Networks in Blocks
Sakana AI proposed DiffusionBlocks—a method that converts residual networks into independently trainable blocks. The idea: interpret layer updates as reverse de
◐ Listen to article
Sakana AI proposed DiffusionBlocks—a method that converts residual networks into independently trainable blocks. The idea: interpret layer updates as reverse denoising steps. This simplifies parallel training and reduces synchronization requirements.