autodiff is cool and all, but have you tried manually computing the gradient to avoid too many variables on the autodiff tape..??? (screenshot of me trying to avoid autodiffing through large Gram matrices)

Playing around plotting some chemical structure embeddings coming from a VAE, and colouring by local intrinsic dimension looks almost like these colour blind test images.

It can be bounded from below and above, by \sqrt{n} and n, respectively. These cases each reflect say an extreme lengthscale, ell=\infty or ell=0, and so it appears almost like a measure of complexity. I'm struggling to come up with an intuitive explanation.

Does anyone have any ideas/references?

Working on a model involving some GP regression, coupled with a horseshoe prior for variable selection. In an attempt at counting the number of effective parameters of my model, the following quantity pops out quite naturally. Is this a well known quantity? 1/2

Klassisk økonomfjas. Det er ikke modellene våre eller måleinstrumentene det er noe galt med — folk lider av «fantomfattigdom» borsen.dagbladet.no/nyheter/nord...

Always fun little tidbits on the wikis of famous mathematicians. Here on Riesz Frigyes' (*of the* Riesz representation theorem) lecturing style

Hmm, from an initial check __not really__. I initially only fit a single chain, so figured maybe I was exploring a single mode, but when I reran the model with 4 chains I still don't really see much multimodality. Maybe n is so small that the prior dominates?

Like, I get that I shouldn't, but it is surprisingly easy to do inference for a very basic BNN in Stan. Just standard NUTS. The sampler complains a lot, but intervals and mean look sensible enough

I love these clear, crisp, cold winter mornings. As long as there is no snow.

Having a lot of fun implementing various versions of the Zig-Zag sampler these days, keen to see how it performs on some more complex models. Screenshot from a "sticky" implementation in linear regression with a spike-and-slab prior