Azhir Mahmood  🚀 🤖 🧬 🧠 🪐

We wrote about the "what, when, and why?" of t-processes: http://proceedings.mlr.press/v33/shah14.pdf . Implementation isn't much more difficult. There are nuanced trade-offs. The uncertainty and dependencies are subtly different and sometimes preferable, the noise model is less interpretable. https://twitter.com/GarridoMerchan/status/1342899907989032961 …

Which situations do you think a t-process would be preferable over GP with Bayesian treatment of scale?

We address the question in depth in the paper. It's actually quite subtle and hard to give a short take. They're mostly useful when you have a small number of observations (e.g. BayesOpt with a really expensive objective). But there is more to the story.