The loss function in score-based (diffusion) models - DiSco MIMO Episode 5

We explained and implemented the loss function for training a score-based (diffusion) generative model. The formula is simple, but understanding it requires us to take a glance at sampling from a diffusion model ahead of time! Paper: https://arxiv.org/abs/2204.07122 Code repository: https://github.com/mariusarvinte/disco-mimo 00:00 Start and recap 00:50 The simplest score matching loss formula 03:22 Explaining the loss function (and diffusion sampling in general) 15:00 Implementing the loss in PyTorch 24:35 Computing the loss with random data and untrained models