Brilliantly Wrong — Alex Rogozhnikov's blog about math, machine learning, programming, physics and biology.
Recently, a pivotal moment in protein design happened: binding affinities and success rates have improved rapidly. Protein design is already a very hot area, and the distance from idea to potential solution has shortened dramatically. Here are some reflections on the importance of this moment and some looking-up thoughts for the area. Antibodies for the next epidemic Let’s roll back in time to…
Last year Pascal Notin wrote a great post summarizing important observation about AI + proteins: Have we hit the scaling wall for protein language models? . (Spoiler: the answer is ‘yes’) Briefest summary if you didn’t read it: PLMs’ performance on fitness prediction (‘transferability’ of skills) plateaus after 1B and declines after 5B parameters. This holds for multiple PLM families leading…
Who needs fast autograd? Seemingly everyone these days! And once upon a time I needed an autograd that is actually fast . Leaving project details aside, here are the requirements: we test many computation graphs (graph is changing constantly) many-many scalar operations with roughly 10k—100k nodes in each graph every graph should be compiled and ran around 10k times both forward and backward this…
This month brought two preprints describing optical pooled CRISPR screens. What’s this new technology, what it can be used for, and why I’ve been waiting for it? I’ll make a small comparison of approaches and critically review the papers. Best of all — I am not affiliated with either team, and this is likely the most unbiased review you’ll find 😅 Papers discussed: PERISCOPE aka Perturbation…
Einops is soon-to-turn 5 years. Right time to have a look back. Some intro: einops is widely used — around 4 million downloads a month (for calibration - pytorch is 10 million) on pypi and is used in thousands of projects on github. In a number of ways einops is unique: bends tensors for a number of very different frameworks. AFAIK all other efforts to make something truly multi-framework either…
A great achievement of the past decade in programming is a shift in paradigm from transition -focused to state -focused. This shift is clearly seen in front-end (user interfaces): In react/preact/vue and other frontend frameworks a component has a state and defines how state should be represented (rendered) in html. The aim of a framework is to ‘migrate DOM’ to desired html representation with…
Summary I argue for wider usage of delimiter-first in the code three friends [tic, tac, toe] becomes three friends ・tic ・tac ・toe . A new top-level syntax for programming languages is proposed to show advantages of this method. New syntax is arguably as simple, but more consistent, better preserves visual structure and solves some issues in code formatting. Related: comma-first formatting A…
If you want to learn some culprits of microscopy … you’d better watch this video by microbehunter, because rest of the post is view of ML person on things you should (not) expect from lab microscopy during experiment design. Warning: This post contains reflections and is not meant to be an easy reading. This post assumes that you understand wave mechanics. I have a nice general background in…
(a friendly reminder that reading post before commenting is a great idea. Some people see this as an argument for GUI, but it is completely misleading) A favourite activity of fresh github-bers is writing CLI (command-line interfaces) for anything. Every programmer uses CLI (true) , so writing CLI makes you more professional (false) . CLIs are required in everyday maintenance, env/pipeline/db…
Abstract: Frequently comparing deep learning models? A simple way to improve comparison is discussed here, this trick becomes specially handy when comparing segmentation models. Reliable comparison of models is a question important for DL “theorists” (to evaluate new approaches) as well as for practitioners/engineers (to select an approach for a particular task in hand). Comparison is…