Closed aleximmer closed 3 days ago
For sequential decision making (BO & contextual bandits), there doesn't seem to be any difference between post-hoc and online marglik, in my experience. See
I also observed this in my latest BO-with-LLM paper, although the results are not published.
I think it's a matter of putting this as a caveat in the README.
Makes sense. I think it's just a matter of generally making sure that people don't expect everything to work automatically and be optimal but rather try different options depending on the problem.
In contrast to the redux paper, we give an example with post-hoc Laplace and marginal likelihood prior precision tuning in the docs (https://aleximmer.github.io/Laplace/#full-example-post-hoc-laplace-on-a-large-image-classifier). This is confusing and we have not thoroughly tested this approach but only recommended full + online marglik or last-layer + validation-based prior tuning.