Twitter | Search | |
Sara Hooker
Research @ Google Brain, interested in model compression, robustness + interpretability. Founder of data for good non-profit .
1,590
Tweets
1,231
Following
8,328
Followers
Tweets
Sara Hooker 8h
Replying to @AndrewLBeam @hardmaru
I do think the debate has centered on a narrow enough topic -- bayesian neural networks -- that it actually feels less about "what tribe you are part of" and more anchored. That's probably why it also feels like a welcome break from the debate about the scope of deep learning :)
Reply Retweet Like
Sara Hooker 8h
Replying to @hardmaru
Agreed. :) Far more fun.
Reply Retweet Like
Sara Hooker retweeted
hardmaru 8h
I’m happy to see ML Twitter debating things like the pros and cons of Bayesian approaches, rather than things like defining the scope of Deep Learning, or TF vs PyTorch.
Reply Retweet Like
Sara Hooker retweeted
iandanforth 17h
Bayesian language is so obfuscating. If I said my first "guess" doesn't matter, or that my "hypothesis" doesn't matter, it would sound absurd, but call it a "prior" and people start nodding along ...
Reply Retweet Like
Sara Hooker 14h
Replying to @carlesgelada
100% agree. the personal attacks on twitter in dec towards indicate unwillingness to engage with the debate itself and also discourage researchers from openly disagreeing or exploring scientific ideas. The community needs to improve conduct on forums like twitter
Reply Retweet Like
Sara Hooker 14h
Replying to @sarahookr
I am also glad carles + jacob reflect upon the december twitter discourse at the end of the blog "We all understand that putting your ideas out in the world means that they will get critiqued and picked apart... but the discussion should stay exclusively on the science. "
Reply Retweet Like
Sara Hooker 14h
I really enjoyed reading the post by + both because 1) it prodded me to re-examine my own opinion on BNNs, and 2) because questioning the assumptions that appear to be set in stone is healthy for the research community. :)
Reply Retweet Like
Sara Hooker 14h
This post follows on the heels of "the case for Bayesian deep learning" by + a heated twitter debate last december which was ultimately disappointing because of fairly personal attacks on by more senior researchers.
Reply Retweet Like
Sara Hooker 14h
Back to back interesting blog posts - "...when a Bayesian tells you that BNNs provide good uncertainty estimates.. We should ask, “what evidence are you providing that your priors are any good?” New blog post by +
Reply Retweet Like
Sara Hooker Jan 16
Replying to @ChombaBupe
Do you have a write up of this somewhere ?
Reply Retweet Like
Sara Hooker retweeted
Andreas Madsen Jan 16
After my blog post, many professors have asked me what the alternative is to hard-filtering by "1–2 publications in top ML venues". – My pragmatic opinion is: "just be honest, mention it on your website". Then students are empowered to decide whom they will spend their time on.
Reply Retweet Like
Sara Hooker retweeted
DataScienceNigeria Jan 16
World class research from on a formal methodology to evaluate the impact of pruning Neural Networks at class & exemplar level.Results shed light on previously unknown trade-offs & suggest caution on the use of pruned models in sensitive domains:
Reply Retweet Like
Sara Hooker Jan 15
Replying to @JeffDean
I found reading reading the history + physics fascinating the most interesting human element of the story is that Fosbury initially experimented with it because he couldn’t compete using existing techniques — his “crazy” method was out of necessity :)
Reply Retweet Like
Sara Hooker Jan 15
Replying to @JeffDean
Different sport but I recently found out about the history of the “back-first” technique for high jump. I couldn’t believe it was introduced as recently as the 1968 olympics. It took a trailblazer to try it because It feels unnatural,but it turns out it lowers the center of mass
Reply Retweet Like
Sara Hooker retweeted
Jeff Dean Jan 15
I think I'd never really watched a pole vault in such slow motion from this angle. It truly is an incredibly complex motion!
Reply Retweet Like
Sara Hooker Jan 15
Replying to @_dsevero
Yes! This direction of research — to use adversarial techniques to improve fairness outcomes is very interesting to me.
Reply Retweet Like
Sara Hooker Jan 15
Replying to @sarahookr
Btw... I have said this before, but this is why facial recognition should absolutely not be used in a sensitive domain like hiring at all.
Reply Retweet Like
Sara Hooker Jan 15
“Adversarial” attacks are often perceived as bad, but if the model itself is problematic adversarial solutions like “how to compose your face” can be a way to try and address biased feature extraction.
Reply Retweet Like
Sara Hooker Jan 14
this gave me chuckles why don't you want a free trip to the moon with guaranteed safe return? "nothing/not enough to see or do" what is this mad world we live in.
Reply Retweet Like
Sara Hooker retweeted
Drew Moxon Jan 13
I open Twitter to have my mind blown by "reconfigurable organisms." Not only is the technology itself absolutely amazing, but the frontier of what it could bring next creates so many more questions, both exciting and scary.
Reply Retweet Like