should we be surprised at this point?

🌎

Of all sad words of LLM, The saddest are these: "Failed to fetch arxiv.org" again

a metaphor for x (formerly twitter) @norvid-studies.bsky.social

We've been investing heavily in better protein language models (PLMs), but relatively little work addresses how to best generate with them. We present a new search-based method for PLMs and exhaustively benchmark models and methods, including with in vitro data from antibody therapeutics campaigns.🧵

The relationship between Spearman correlation and AUROC computed on the same data (one continuous variable, one binary variable) is surprising. On evenly balanced data, Spearman=0.75 leads to AUROC=0.93. As the binary variable gets more imbalanced, the discrepancy gets bigger.

The Mann-Whitney U statistic and rank-biserial correlation are simply rescalings of AUROC: U = AUC * n_pos * n_neg r_rb = 2 * AUC - 1

What makes tabular data unique (and interesting!) is not merely that it's arranged into rows and columns. New blogpost: calvinmccarter.substack.com/p/the-idiosy...

Has anyone tried far-UVC in their home? It's now dropped into the ~$300 price range where I'm interested in trying it for myself. substack.com/home/post/p-...

Here's a link to the report: cleanlabelproject.org/wp-content/u... (TLDR heuristics: whey is better than plant-based, non-organic is better than organic, unflavored is better than chocolate-flavored)

Ultra exciting! And it's gratifying to see that this method uses the kernel density integral preprocessing method that I published in @tmlr-pub.bsky.social (2023). (One takeaway: even if your ML research focus isn't deep learning, pursue directions that complement rather than compete with it.)

How do you go from a hit in your antibody screen to a suitable drug? Now introducing CloneBO: we optimize antibodies in the lab by teaching a generative model how we optimize them in our bodies! w/ Nat Gruver, Yilun Kuang, Lily Li, @andrewgwils.bsky.social and the team at Big Hat! 1/7

Will neural networks achieve AGI before they figure out how to do tokenization internally? Or, on the way to AGI, will they invent a tokenizer that "just works"? Another way of framing it: is tokenization AGI-complete?

JMLR and TMLR provide better reviewing, but worse publicity for accepted papers. Their only real platform, their accounts on X, get very little engagement these days. They should be on 🦋. Better yet, *MLR should have a biweekly or monthly arXiv-style email newsletter. @thegautamkamath.bsky.social

Pretty interesting new paper in TMLR: "Controlling the Fidelity and Diversity of Deep Generative Models via Pseudo Density" openreview.net/forum?id=8Vk...

Twitter is still a better place to share ML papers than Bluesky:

Excited to share a new paper published in TMLR, "Towards Backwards-Compatible Data with Confounded Domain Adaptation", solving a key problem in using AI for biology. How can you combine datasets across settings, when "what you measure" and "how you measure it" are confounded? 1/7

Suppose you collect tissue samples with varying degrees of freshness, and want to correct for this. But what if freshness is also correlated with true biological differences (eg healthy vs cancer tissue)? My new paper out in TMLR addresses this problem: openreview.net/forum?id=GSp...