I don't think the official Sakana account has posted this here yet. DiffusionBlocks was a recent ICLR paper that showed you can divide networks into blocks and train each independently, reducing the memory required for training. pub.sakana.ai/diffusionblo...
pub.sakana.aiDiffusionBlocks: Training Neural Networks One Block at a TimeA principled framework that converts a residual network into independently trainable blocks via a diffusion interpretation, achieving B× memory reduction without sacrificing performance.
3 reposts15 likes