I resigned from Meta, and my internal post got leaked to the press, resulting in some fragmented quotes. Here is the full thing: https://t.co/iUcr8TYMLD
Super excited by this new #ICML paper: we find that many deep nets have a few (<50) "critical" neurons and we have an efficient alg to find them.
Evaluating these critical neurons enables fast model repair (e.g. reduce bias) and interpretation. Great work by @GhorbaniAmirata!
@hippopedoid@jhnhw I occasionally encountered similar artifact with scRNA seq data. The spaghetti-loop in UMAPs changes completely when you shuffle the order of rows in the training data.
The graphics in this article are excellent— this is a huge part of communications that has been missing. If you want to understand how layered interventions reduce #covid19 spread— why masks are indispensable but not always “enough” — read this.
https://t.co/r1KfeKHW5J
1/5 Self-Distillation loop (feeding predictions as new target values & retraining) improves test accuracy. But why? We show it induces a regularization that progressively limits # of basis functions used to represent the solution. https://t.co/570qXFmlGj w/@farajtabar P.Bartlett
Trolley problems ("Kill 1 to save 5?") are good for clarifying moral intuitions, but my grad student Julian DeFreitas argues they're irrelevant to designing driverless cars, which should kill no one. https://t.co/9kXzHhsKOO
My biggest regret as a 1st year grad student was pretending to say I knew X when I didn't really. Learning to say "No, tell me about X" became a gateway for (1) learning stuff, (2) getting to know a person, and (3) learning how to say no in other areas of my life.
Some news! After many lovely years at @harvardmed I'm moving to @Columbia fall 2020 to start a new lab as an Assistant Professor in Systems Biology and the Program for Mathematical Genomics--and I'm recruiting students and postdocs! Email/DM me or see https://t.co/x3SHeZPFiD. 1/2
Interested in using deep learning for designing proteins or cis-regulatory sequences? Try out this new (and fast) method by the very talented @jjohli: https://t.co/1wfb912Bgz
Join me at 8PM ET tonight for a live @munkdebate Q&A, where we'll discuss #covid19 and the post-pandemic world.
Tune in on Facebook: https://t.co/6MYgYDvkHZ
Or on the Munk Debates website: https://t.co/uzlJlGoaUL
This is a really neat way of using neural networks to get rid of polygonal meshes. The interaction of computer graphics with neural networks is really exciting.
(1/2) New study just out, face masks effective for source control against influenza and seasonal coronaviruses, even reducing virus detection in aerosols in patients with coronavirus infections. 5 years of hard work to generate the data https://t.co/nXJOKLVTvk
More evidence of pre-symptomatic transmission of COVID19, which I think we must now regard as the reality we are facing. This paper estimates that 44% of transmission could occur 2-3 days before first symptoms. https://t.co/6mkH9AvtkL via @marcelsalathe#quarantineloophole 30/