Day 31: TPU-PyTorch Support !?

Multi-Modal Methods: Visual Speech Recognition (Lip Reading) —- hundreds of thousands of people with speech impairments, every year, could one day benefit from this technology https://t.co/lZMN6a62UW

— Nando de Freitas (@NandoDF) June 19, 2018

Best paper award at #CVPR2018 main idea: study twenty five different visual tasks to understand how & when transfer learning works from one task to another, reducing demand for labelled data.

Paper: https://t.co/18HUu6QnHx
Data: https://t.co/HIBR1mR9yq pic.twitter.com/Ya0jAGh5pp

— Reza Zadeh (@Reza_Zadeh) June 19, 2018

Congratulations to Best Paper Award by @zamir_ar @silviocinguetta @StanfordCVGL & collaborators at #CVPR2018, a computational map of perceptual task transfer learning!! https://t.co/1pLD4hVwmO

— Fei-Fei Li (@drfeifei) June 19, 2018

GraphRNN: Generating Realistic Graphs with Deep Auto-regressive Model by @youjiaxuan, R. Ying, @xiangrenUSC, @williamleif, and @jure#ICML2018 Paper: https://t.co/OsM0cIAH1o@PyTorch Code: https://t.co/Z8AiTSCOIt
Unofficial Slides: https://t.co/hz6sWugVBn pic.twitter.com/8cg7JobrX2

— Brandon Amos (@brandondamos) June 20, 2018

Google Cloud TPUs now offer preemptible pricing at ~70% off the reserved instance pricing. This means, for example, that you can train a ResNet-50 model for ~$7.50 instead of $25, or a Transformer neural translation model for ~$13 instead of $41.

See:https://t.co/EnJXQjbuhN pic.twitter.com/2fBmY3P6Bf

— Jeff Dean (@JeffDean) June 19, 2018

Spot-pricing on TPUs is getting good. We have prototyped TPU-PyTorch support (w. Google engineers), hammering coverage and performance now. Promising times.... https://t.co/v7Lz6I2JWH

— Soumith Chintala (@soumithchintala) June 19, 2018

DensePose is now open source (and at the top of Hacker News yesterday). https://t.co/uJ0pwkgw5P

— Yann LeCun (@ylecun) June 20, 2018

In a letter to @satyanadella, Microsoft employees are asking the company to terminate its contract with ICE https://t.co/ZVParowmr5

— kate conger (@kateconger) June 19, 2018

Microsoft workers request MS:

1. Cancel the existing Azure Gov contract with ICE immediately.

2. Draft, publicize, enforce clear policy stating that neither MS nor its contractors will work w/clients who violate int'l human rights law.

3. Commit to transparency and review... https://t.co/gAHL0bAbUm

— Lilly Irani (@gleemie) June 20, 2018

Tech employees should question what applications their tech is being used for. Employees have more power now as tech companies have evolved to be run by engineers and scientists over the past decade. https://t.co/DxpQqjyMDc

— hardmaru (@hardmaru) June 19, 2018

Here’s @satyanadella echoing our earlier reporting, saying Microsoft’s ICE contract doesn’t involve facial recognition. No explanation for the January blog that claimed otherwise https://t.co/4Puv7629fD

— kate conger (@kateconger) June 20, 2018

Interesting: https://t.co/VxldDWw8h1 "modern" CNN classifiers, using less pooling layers (compensated by strided conv), are more accurate than older ones, but at the same time less invariant to translation because of Shannon's theorem @rasbt @fhuszar @roydanroy @TacoCohen

— andrea panizza (@unsorsodicorda) June 19, 2018

"Deep Neural Decision Trees," Yang et al.: https://t.co/655f9uBWQz

— Miles Brundage (@Miles_Brundage) June 20, 2018

Horizon charts of realtime data. https://t.co/SpBlpdwDCn

— Mike Bostock (@mbostock) June 19, 2018

What is the role of qualitative methods in addressing issues of replicability, reproducibility, and rigor? https://t.co/aJ33RXF8wx

— Andrew Gelman (@StatModeling) June 19, 2018

"Deep Learning for Dialogue Systems" - great [TUTORIAL] by @YunNungChen https://t.co/Dq8xmHriLe

— Xavier 🎗🤖🏃 (@xamat) June 20, 2018

Monte Carlo Tree Search - beginners tutorial https://t.co/Na5k8mhrJv

— Nando de Freitas (@NandoDF) June 19, 2018

Apex is a PyTorch extension from @nvidia that makes it easy to use mixed-precision training and use Volta Tensor Cores to full potential. Read more at: https://t.co/9xuOLAyGJT

— PyTorch (@PyTorch) June 19, 2018

How might visualizations help people reason more cautiously in the face of uncertainty? Value-Suppressing Uncertainty Palettes intentionally make uncertain values harder to distinguish! https://t.co/TciyeqjSHs

— Interactive Data Lab (@uwdata) June 19, 2018

I love do 🖤 a good workflow…
"File organization best practices" by @abtran https://t.co/PE7lLypdkX #rstats pic.twitter.com/ZvRX244aAr

— Mara Averick (@dataandme) June 19, 2018

ICYMI, 🎞 step-by-step monte carlo w/ magick!
"Animating a Monte Carlo Simulation" by Thomas Rohhttps://t.co/YOCtIYLZBM #rstats #dataviz #infovis pic.twitter.com/XZJsXI2YNS

— Mara Averick (@dataandme) June 19, 2018

Remember engineers, have ethics. If you are asked to do something illegal or morally wrong then say no. You can go to jail for doing what the boss says. https://t.co/wY7Hh51Y80

— Jim Bennett ☁️ (@jimbobbennett) June 19, 2018

Kjell and I are writing another book on predictive modeling, this time focused on all the things that you can do with predictors. It's about 60% done and we'd love to get feedback. Check it out at https://t.co/sU9nyCcSql #rstats #MachineLearning pic.twitter.com/T1tVcZAmcL

— Max Kuhn (@topepos) May 14, 2018

Deep Learning; The study of neural networks and the art of making all kinds of stuff differentiable (including things that really shouldn’t be)

— Denny Britz (@dennybritz) June 19, 2018

Official launch of the competition for software developers to use graduate salary & employment outcomes data to create apps to shape prospective students' choices about universities & degrees - open to 'an organisation of any size'... https://t.co/20K2DS6hL0

— Ben Williamson (@BenPatrickWill) June 19, 2018

Surprise! Patent trolls don't lead to innovation or knowledge sharing. Largest survey yet. https://t.co/cIOCJ49X7c

— James Bessen (@JamesBessen) June 19, 2018

An important lesson that I wish I could impart on everybody: A metric is usually a proxy. Getting very good at the metric means it is time to perform qualitative analysis whether the metric achieves the qualitative improvements it was designed to proxy for.

— halvarflake (@halvarflake) June 19, 2018