Our #icml2020 paper "Involutive MCMC: a Unifying Framework" is now available on arxiv arxiv.org/abs/2006.16653. It describes many MCMC algorithms from a single perspective. Work with Max Welling, Evgenii Egorov, Dmitry Vetrov

Finally someone scaled up VAEs! Since Glow (and autoregressive models) we knew likelihood-based methods can generate realistic images at the expense of requiring lots of parameters, so the same should have been possible for VAEs. And now we know it is!

Here's a condensed version of the matplotlib cheatsheets so it can fit a desktop background (github.com/matplotlib/che…) Full image: drive.google.com/file/d/1kwYFaR… and vectorized .svg, with the non-standard fonts outlined: drive.google.com/file/d/1b2LtZU… Thanks Nicolas P. Rougier et al for making it!

KL divergence asymmetry

Really liked this talk by Kilian Weinberger at ML-RSA workshop #NeurIPS2020 on why just beating the SotA isn't something one should strive for and to keep digging until a simple explanation of what makes it a SotA is found. slideslive.com/38938218/the-i…

We have this awesome function called sublots_mosaic where you can pass us a layout id'ed on name axd = plt.subplot_mosaic( """ ABD CCD """) matplotlib.org/stable/tutoria…

Tired of Jensen's inequality working in a wrong direction? Try reciprocal convexity! artem.sobolev.name/posts/2021-05-…

Come by our poster "On the Periodic Behavior of Neural Network Training with Batch Normalization and Weight Decay" tomorrow at #NeurIPS2021 poster session 6! With Maxim Kodryan, Nadia Chirkova, Andrey Malinin, and Dmitry Vetrov. Poster: nips.cc/virtual/2021/p…

Sad to see Twitter going through a lot of turbulence these days. If you're on mastodon follow me @[email protected]

You can play chess with #ChatGPT, but it appears it's not very good at it. Curious to see how much better it could get with simple fine-tuning on a large base of chess plays.

Just stumbled upon this gem by John Schulman – a very interesting read on estimating KL divergences in practice and possible bias-variance tradeoffs joschu.net/blog/kl-approx…

Best paper I've read so far this month: All elementary functions (sin, cos, tan, exp, log, powers, roots, hyperbolic functions, π, e, and even basic arithmetic) can be generated from just one binary operator: eml(x, y) = exp(x) − ln(y) …plus the constant 1.