This site now comes with archive pages that make history posts more accessible. Next on the list is adding a search box.

Unbalanced Training Data

I just published “Deep learning unbalanced training data?Solve it like this.” https://t.co/PhDNgsnaDv

— shubrashankh (@shub777) May 26, 2018

What to do when there are only 1-2 images per class in image classification. I used the approach which @jeremyphoward mentioned in one of his deep learning1 lectures(Just copy the unbalanced class). Simple and effective. #DeepLearning #unbalancedclass https://t.co/YJpGvUhaf4

— shubrashankh (@shub777) May 26, 2018

Amazingly fast and simple top 10% Kaggle solution https://t.co/CEAcaqiMS2

— Jeremy Howard (@jeremyphoward) May 27, 2018

On ACL Pre-print Policy

(Part of a long thread. Please click on the tweet to read the full thread.)

I saw another thread critical of the @ACL_NLP preprint policy (with @kchonyc @haldaume3 @lintool @emilymbender @thedansimonson). I think the ideas underlying the policy have been insufficiently explained to people, so here’s one more attempt. 1/

— Christopher Manning (@chrmanning) May 27, 2018

botcheck.me

Identifying Propaganda Bots on Twitter - Methodology

Check out https://t.co/iLfE1jzZ1g to analyze and recognize twitter bots for propaganda!

— KDnuggets (@kdnuggets) May 26, 2018

Notables

More machine learning! We found that it works great in combination with unsupervised timeseries segmentation (@gordonberman's MotionMapper or @awiltsch's MoSeq) to cluster out the body part trajectories into distinct "movemes" pic.twitter.com/bsuhIZKPml

— Talmo Pereira (@talmop) May 25, 2018

The suggested paper on interval estimation for binomial proportions is a great read: https://t.co/0OgDCPLIkk https://t.co/DNlBHE24bI

— John Myles White (@johnmyleswhite) May 26, 2018

Miscellaneous

dlib 19.13 is now out. Added about 50 new classes and functions to the Python interface: https://t.co/YpXr6h04xg

— Davis King (@nulhom) May 26, 2018

Scientific and societal impact of research comes rarely from where we expect. Open source software is one such example: today, Scikit-Learn just passed 10000 citations! Congratulations to all contributors for all the progress they triggered! pic.twitter.com/dkhzbVKI6L

— Gilles Louppe (@glouppe) May 26, 2018

"These dynamic langs are so sloppy. We should be more rigorous, like maths."
"Cool! What does maths use to indicate types?"
"Fonts, mostly."

— Chris Ford (@ctford) July 18, 2017

not a huge fan of the term "differentiable programming". The big deal isn't that it's differentiable, it's that there is any optimization over the code at all, instead of explicit code. I like/started to use "fill-in-the-blanks programming", artifacts of which are Software 2.0 :)

— Andrej Karpathy (@karpathy) May 26, 2018

Fun 50-minute documentary about Canada's role in the latest developments in AI, focusing in part on Geoff Hinton, Yoshua Bengio and Richard S. Sutton.
Produced by Bloomberg and hosted by... https://t.co/Crgneamfy7

— Yann LeCun (@ylecun) May 26, 2018

I get messages a lot that start with "Have you ever thought of..." or "Did you consider..." Trust me, I've heard your idea before. The world isn't lacking ideas, it's lacking good execution. If you believe in your idea, go out and try to make it successful.

— comma ai (@comma_ai) May 25, 2018

I think a lot of "best practices" arguments in programming are more about social power and in / out group dynamics than anything evidence based. Same goes for framework arguments, which language to use arguments etc.

— taotetek (@taotetek) May 26, 2018

There are no Tony Starks. There are only assholes who take everyone else’s ideas, put them together and claim they are Tony Stark.” -@doctorow #phoenixcomicfest2018

— Michael Senft (@RelentlessRead) May 26, 2018

Response to "But all my nearest neighbors are doing it!!" https://t.co/6KC3VzsZ1o

— Data Science Renee (@BecomingDataSci) May 27, 2018

@ceshine_en

Inpired by @WTFJHT