3 ms·
Honestly, as a mathematician-turned-ml-researcher, Song and Ermon’s mathematical derivations based on SDEs are what sold me on diffusion/score-based generative
by catgary 3y ago
Honestly, as a mathematician-turned-ml-researcher, Song and Ermon’s mathematical derivations based on SDEs are what sold me on diffusion/score-based generative models. This (https://youtu.be/wMmqCMwuM2Q?si=fvujznGDHjKH2yUi https://youtu.be/wMmqCMwuM2Q?si=fvujznGDHjKH2yUi) is probably the best ML seminar talk I’ve ever seen.
- samsartor 3y agoOh definitely. I think that 2021 paper was what sold me as well! But I don't think it is very approachable to non-technical audiences, and there are a number of other was to interpret diffusion models which are pretty intuitive. That is also why I like the alpha-(de)blending paper from this year.