Day 37: OpenAI Five Defeats Amateur Human Teams

New NLP Newsletter w/ even more cool stuff this time! Transfer learning, Chris Olah, Software 2.0, NMT with attention notebook, gradient boosting in-depth, Defense Against the Dark Arts, interpretability and bias, RL, scene understanding https://t.co/qo0y677c0F

— Sebastian Ruder (@seb_ruder) June 25, 2018

OpenAI demonstrates remarkable progress in a limited version of 5v5 Dota using two concepts that we didn't think can learn long time-scale strategies: selfplay, LSTM. Carefully designed reward functions are notable -- intermediate, global, team-spirit.https://t.co/GBTw1e7ERR

— Soumith Chintala (@soumithchintala) June 25, 2018

Amazing what a single-layer 1024-unit LSTM can be trained to do with a bit of engineering! OpenAI Five Model Architecture: pic.twitter.com/mRbD02KpNc

— hardmaru (@hardmaru) June 25, 2018

Some megascale RL results from @OpenAI:
We've scaled existing methods to train AIs with sufficient teamwork skills to solve hard problems within Dota 2
- Scaled-up PPO+LSTM
~120,000 CPUs + 256 GPUs
- Self-play
- Hyperparameter called "Team Spirit" to teach AIs to collaborate https://t.co/lcSGWw0yr5

— Jack Clark (@jackclarkSF) June 25, 2018

The deep learning moment of deep RL: https://t.co/25qB39K3HL

— Ilya Sutskever (@ilyasut) June 25, 2018

Great work by @OpenAI team. More evidence that scaling up simple RL methods (rather than designing complicated algorithms) enables solving increasingly complex problems. https://t.co/1wpazo69hU

— Arthur Juliani (@awjuliani) June 25, 2018

I want to wax poetic about the models (LSTM+PPO pushed far beyond what people likely thought possible, mirroring @Smerity et al + @GaborMelis et al in LSTM language modeling), the game (DotA as a complex test-bed), or the stupendous compute (180 years of gaming per day), but ... pic.twitter.com/cgUJ4w39G0

— Smerity (@Smerity) June 25, 2018

Official version out! Did you know that craving actually increases, rather than declines, during early abstinence? Here's why - https://t.co/UwgyHHKv4G#addiction #addictionrecovery

— Dr. Xiaosi Gu (@xiaosigu) June 25, 2018

Dense Object Nets: Learning Dense Visual Object Descriptors By and For Robotic Manipulation. Check out those sexy descriptor images. https://t.co/3b5mL2juHz #computervision #robotics pic.twitter.com/r9pFEoQpK4

— Tomasz Malisiewicz (@quantombone) June 25, 2018

We present 𝗥𝗲𝗻𝗱𝗲𝗿𝗡𝗲𝘁, a CNN for differentiable rendering of 3D shapes. RenderNet can perform both rendering (for different shaders with the same network), and inverse rendering (for tasks like single image reconstruction and novel view synthesis) https://t.co/9Muy1vTEPY pic.twitter.com/QlUMumT6Z9

— Thu Nguyen Phuoc (@thunguyenphuoc) June 25, 2018

Finished 4th out of 155 participants at the @udacity @lyft Perception Challenge with the fastest pipeline.
The task was nearly real-time semantic segmentation of images from a front facing camera (car and road classes).
Details: https://t.co/ffgOcSujEl#udacitylyftchallenge pic.twitter.com/moJZ5yS2os

— Nick Falaleev (@NFalaleev) June 11, 2018

If you are using PCA to visualize neural network training trajectories, you are interpreting it wrong! Very proud of this work with @joe_antognini: "PCA of high dimensional random walks with comparison to neural network training" https://t.co/04I1D8vtJl pic.twitter.com/i9vSfz9ZaR

— Jascha (@jaschasd) June 26, 2018

Lots of research on neural net pruning/distillation/etc. but is any of this stuff used in practice? If it works as well as people often claim, I'd expect it to be used regularly for faster inference at big tech companies, but don't think I've seen any claims to this effect.

— Miles Brundage (@Miles_Brundage) June 26, 2018

#Map compares #USA climate with equivalent cities from around the world. Source: https://t.co/3KimQijJDj pic.twitter.com/OJqRxfMFlX

— Simon Kuestenmacher (@simongerman600) October 6, 2017

Check this out:https://t.co/0bxsgTeZNJ pic.twitter.com/1Ikg8utFiU

— Micah Cohen (@micahcohen) June 25, 2018

📊 survey results = fun dataset!
"2018 Data Visualization Survey Results" ✍️ @Elijah_Meeks https://t.co/teS3GhHV0F #dataviz #infovis pic.twitter.com/gx7etodpKM

— Mara Averick (@dataandme) June 25, 2018

This NMT with attention colab notebook is one of the cleanest and best documented TF examples I’ve seen: https://t.co/eUzJ4aSthD

— Denny Britz (@dennybritz) June 26, 2018

I just published a tutorial on how to train a LSTM network with @HelloPaperspace and sample the resulting model in @ml5js. More tutorials soon! (hint: style transfer and pix2pix)https://t.co/PecQBb51EC

— Cris Valenzuela (@c_valenzuelab) June 25, 2018

"On Intrinsic Rewards & Continual Learning," Satinder Singh: https://t.co/HB4eEyHyTb

— Miles Brundage (@Miles_Brundage) June 25, 2018

🤖 feat. @h2oai, caret + more…
"Using Machine Learning w/ LIME To Understand Employee Churn" ✏️ @bradleyboehmkehttps://t.co/cQnYato5mG via @bizScienc #rstats pic.twitter.com/ggPQuGXvbX

— Mara Averick (@dataandme) June 26, 2018

tableGrob()/grid.table() in gridExtra is a phenomenal way to add summary stats to ggplot-based plots in #rstats. It's super easy to do, too, with annotation_custom() pic.twitter.com/Jqk5rEjZiq

— Andrew Heiss (@andrewheiss) June 25, 2018

Deep learning for time series forecasting: Predicting sunspot frequency with Keras: https://t.co/qC7fXFp7wR #rstats #rkeras pic.twitter.com/k1z3vI8wOu

— RStudio (@rstudio) June 25, 2018

😻 w/ R code!
“Recreating (more) data visualizations from [ @davidmccandless'] ‘Knowledge is Beautiful” ✍️ @MattOldachhttps://t.co/gIk8ObOvFA #rstats #dataviz pic.twitter.com/z1AnrqxnaV

— Mara Averick (@dataandme) June 25, 2018

tf.keras programmer's guide is out! On the v1.9 docs pre-release. Highly recommended, this is a great way to develop with @TensorFlow.

To try it out you'll need to install the rc: pip install --pre -U tensorflow==1.9.*https://t.co/mf4eZxngxi

— Josh Gordon (@random_forests) June 25, 2018

My interview w/ John Ioannidis @StanfordMed, (video + full transcript),
the contrarian and conscience of biomedical (and other) research https://t.co/S0SAhGkZTq
"Most Research Is Flawed, Let's Fix It"@Medscape includes re-published PREDIMED critique

— Eric Topol (@EricTopol) June 26, 2018

In this blog post, I share how data science is like the Model T or the Boeing 747https://t.co/saWOcyTPjM pic.twitter.com/AliMXMTyiE

— David Robinson (@drob) June 25, 2018

1/ My #CVPR2018 summary/run-down:
1) "De-noising is solved"
2) "Fourier is dead. (maybe)". <-- Me
3) All CVPR's < 2018: "We just trained a DNN to get X and it works great lol".
4) CVPR 2018: "Actually jk we can get X from geometry lol" -__-

— Tarin Ziyaee (@tarinziyaee) June 25, 2018

Just because there's intense speculative interest around a piece of tech doesn't mean there's any value in it. But inversely, just because a field is full of scams and get-rich-quick schemes doesn't mean the tech is worthless. Only a first-principles analysis will tell you that.

— François Chollet (@fchollet) June 25, 2018

The Gates foundation spent almost $600m (not including staff time) on an experiment to see if rigorous teacher evaluation tied to performance management would improve results. It didn't. https://t.co/7vygbdC8iu

— Sam Freedman (@Samfr) June 25, 2018

Every course in #statistics #biostatistics #datascience #machinelearning#rstats should BEGIN with the Quartz Guide to Bad Data by Chris Groskopf @onyxfish at least if your doing nonfiction. Data analysis is like painting a house: 90% is in the preparation https://t.co/yErss4ITUk pic.twitter.com/Gc6e1Jq8Aa

— Edward Tufte (@EdwardTufte) June 25, 2018

Looking forward to testifying tomorrow on AGI. It's important for Silicon Valley to have an open dialogue with policymakers about this incredibly powerful potential technology. https://t.co/OIFhHNjY3F

— Greg Brockman (@gdb) June 25, 2018

The cool thing about 2018 is that if your company uses a linear regression at any time you are now a hot AI company somehow

— Austen Allred (@AustenAllred) June 25, 2018

My thoughts on facial recognition & law enforcement. (Spoiler alert: 🚨 They don’t mix well) https://t.co/xRVwB603L8

— Brian Brackeen (@BrianBrackeen) June 25, 2018

In a stunning op-ed, the CEO of facial recognition company Kairos argues that the technology’s bias and capacity for abuse make it too dangerous for govt use.

Kairos will refuse to sell its technology to governments. Via @BrianBrackeen https://t.co/L1ZZv9OpF8 pic.twitter.com/Sixm2SyA8P

— Matt Cagle (@Matt_Cagle) June 25, 2018