Archive
Review
For being an IAP term offering this class was really good. It is taught by Peter Holderrieth who is on the authors list of the Lipman et al. Flow Matching Guide. It is taught from the ODE/SDE perspective. I think it’s easier to learn flow matching and diffusion from a ODE/SDE view at first, i.e. learn what a flow/diffusion model is, then conditional/marginal probability paths, vector fields, losses, score matching. Afterwards, you can circle back to these historical works and see how nice flow matching is compared to doing a whole bunch of invertible neural networks, simulation, or weird tricks to get tractable Jacobians. I give it a 3/5 in difficulty and 4.5/5 in fun.