I'm incredibly proud that the low compute / low resource AWD-LSTM and QRNN that I helped develop at @SFResearch live on as first class architectures in the @fastdotai community :) https://t.co/nVXplj0L86
— Smerity (@Smerity) September 1, 2019
I'm incredibly proud that the low compute / low resource AWD-LSTM and QRNN that I helped develop at @SFResearch live on as first class architectures in the @fastdotai community :) https://t.co/nVXplj0L86
— Smerity (@Smerity) September 1, 2019
I'm creating a Win10 recovery USB key, following 24 hrs of fun involving drivers (NVIDIA, naturally), corrupted registry files, RAM hiccups, etc. I'm increasingly sympathetic to the idea that civilization can't cope with the complexity of software https://t.co/DrGTNr9XUT
— Jack Clark (@jackclarkSF) September 1, 2019
As a new PI, one of my biggest fears is that one day we will make some kind of embarrassing mistake in our research. It's inevitable. That is why we are committed to open data, code, preprints, and where possible, pre-registration. So the community can help us catch errors early.
— Micah Allen (@micahgallen) September 1, 2019
This is 100% true. However, because interviewers (wrongly) value having algorithms / equations memorized, I created my machine learning flashcards. And literally in every interview I did, I looked better because I could write out an equation from memory. https://t.co/yTRFaHBEzA
— Chris Albon (@chrisalbon) August 31, 2019
The fabulous, multitalented @AnnieDuke and I discuss the mechanics of misinformation and and how it manipulates vulnerabilities of the human mind at @WSJ https://t.co/vm2xwonpDn
— Gary Marcus (@GaryMarcus) August 31, 2019
Thanks to ML Twitter & @scikit_learn team for fascinating debate over default values & leaks b/w modeling/software. Takeaways: 1) Maybe change name to RegularizedLogisticRegression? 2) These problems plague all ML/DL libraries, esp. in deep learning & we must take it srsly (1/2) https://t.co/IAmuBZ9Uov
— Zachary Lipton (@zacharylipton) August 31, 2019
What linear thinking looks like https://t.co/dtIBo6UAIM
— Delip Rao (@deliprao) August 30, 2019
By default, logistic regression in scikit-learn runs w L2 regularization on and defaulting to magic number C=1.0. How many millions of ML/stats/data-mining papers have been written by authors who didn't report (& honestly didn't think they were) using regularization?
— Zachary Lipton (@zacharylipton) August 30, 2019
My advice to young scientists looking for blue ocean: pay attention to the lazy intuition pumps everyone assumes to be obviously true. Historically, there is usually gold there!
— Micah Allen (@micahgallen) August 29, 2019
If you use Twitter data for research and are involved in academic use of the data, you may want to pay attention to some upcoming changes that Twitter is planning to make that may cripple the ability to get data from Twitter's API.
— Jason Baumgartner (@jasonbaumgartne) August 28, 2019
Follow the thread here. This looks bad. https://t.co/kNNGzYiArU
Two good TPU articles just went out:
— Martin Görner (@martin_gorner) August 27, 2019
The secrets of bfloat16, the floating point format TPUs use to run faster: https://t.co/OQp3cByXyR
Code comes from people. It impacts people. The people who are contributing impact who else contributes in the future. If you don't see why accepting a PR from Hitler is about more than code, then you're missing so, so much.
— Mara Averick (@dataandme) August 27, 2019
Technology isn't neutral. https://t.co/OpRlHbDPS8