Day 3: Translate Bach to Beethoven

"A Universal Music Translation Network," Mor et al.: https://t.co/iUzXTZkbH8

Samples: https://t.co/pXAMyx7ydK

— Miles Brundage (@Miles_Brundage) May 22, 2018

"A Universal Music Translation Network"
fun work from my colleages that translates one music organ / style to anotherhttps://t.co/itFge4IUo3https://t.co/TQOvsvodpB

— Soumith Chintala (@soumithchintala) May 23, 2018

Very cool work! 🎵 https://t.co/hPJ0TM5wxD

— hardmaru (@hardmaru) May 22, 2018

The constant barrage of awesome-looking AI results can induce panic in AI researchers. How could you and I possibly keep up, let alone compete? I try to remember the following:
* Most AI researchers oversell their results
* Most “breakthroughs” rely on existing well-known methods

— Julian Togelius (@togelius) May 22, 2018

* ArXiv is not peer-reviewed, and some of what’s pushed out there will not replicate, or only works under very specific circumstances
* Many researchers do atrocious literature reviews, and will claim novelty even when that’s dubious

— Julian Togelius (@togelius) May 22, 2018

* You will be able to read a textbook in five years, summing up all of the AI developments from 2018 they are actually worth caring about - probably in just a few pages

— Julian Togelius (@togelius) May 22, 2018

Perhaps most importantly:
* The world is large, and there’s still room for your research

— Julian Togelius (@togelius) May 22, 2018

You know, Julian, my personal feeling of AI being overhyped has gotten so much worse after I joined Twitter last year.

— Christoph Salge (@ChristophSalge) May 23, 2018

A good thread about the flood of AI preprints. https://t.co/8NhwF00rUT

— François Chollet (@fchollet) May 22, 2018

"A good reminder that Data Science Central is completely insane" https://t.co/68esZTv88d

— mat kelcey (@mat_kelcey) May 22, 2018

I'm blocked, so can't see the tweet, but can imagine it's the typical DSC nonsense. I regularly deter beginners from their unfortuately-easily-found sites, for many many reasons. And I call out prominent tweeters that continue to share their content.

— Data Science Renee (@BecomingDataSci) May 22, 2018

I'm pretty sure it's effectively a keyword farm & they are just after ad clicks. The content is beyond being technically wrong, it's often just reads like random text.

— mat kelcey (@mat_kelcey) May 23, 2018

Amazon is selling its face-recognition tech to police departments https://t.co/X1Y28nnU5A

— MIT Tech Review (@techreview) May 22, 2018

Machine learning could end up being a dangerous threat to our liberties. It's going to be crucial for practitioners to fight so ML doesn't become a tool for oppression. https://t.co/OgSShdBPVF

— Bharath Ramsundar (@rbhar90) May 22, 2018

The FBI:
- has access to 411 million facial images
- if true candidate exists in database, will be listed by their system in top 50 choices just 85% of the time
- claims they don't need to worry about false positives
from @EFF report https://t.co/SMu3OI8H6M pic.twitter.com/wYti4tKfLv

— Rachel Thomas (@math_rachel) May 22, 2018

🙋‍♀️ Are you curious about using Machine Learning in the browser? I took a basic TensorFlow.js example, put a neat graph on it and commented every line of code, so you can play with it!!

💥 https://t.co/WQvrRZ7b4Q pic.twitter.com/UzBWPu3cU0

— Monica Dinculescu (@notwaldorf) May 22, 2018

Last year I said superhuman poker AIs would be running on smartphones in 5 years. That timeline may have been pessimistic. Our new paper "Depth-Limited Solving for Imperfect-Information Games" shows how to develop a master poker AI on a 4-core CPU: https://t.co/xzY4bc1blr pic.twitter.com/oXWJt4hgzm

— Noam Brown (@polynoamial) May 22, 2018

Detailed writeup of how Airbnb uses deep learning to classify our 170m+ listing photos and more. Nice writeup Shijing and Qiang! https://t.co/sbYjdJe0vB

— Mike Curtis (@mikecurtis) May 22, 2018

Learn NLP the smart way, from the best.

AAN ("All About NLP"), the search engine for NLP tutorials, is out now, after years of work: https://t.co/UFc6mYgZTo

ACL paper: https://t.co/1uHQjJ4t24

Blog post: https://t.co/IaIhm4X6UL

LILY group: https://t.co/5dbfXUyy0A pic.twitter.com/pldKFhWHOn

— Dragomir Radev (@radevd) May 22, 2018

How I build an AI to play Dino Run via @Acing_AI https://t.co/tv0xUSteo3 #Python #ArtificialIntelligence #MachineLearning pic.twitter.com/dGxtXnU3Ur

— Python Weekly (@PythonWeekly) May 22, 2018

This neural net will generate a soundscape to go with wherever in Google Maps you happen to be viewing. https://t.co/ay06VGMQn5
Top-notch demo by @qosmo_inc

— Janelle Shane (@JanelleCShane) May 22, 2018

Self-attention for GANs. No more problems with losing track of how many faces the generator has drawn on the dog. https://t.co/WOWglD3ft5

— Ian Goodfellow (@goodfellow_ian) May 23, 2018

This paper shows how to make adversarial examples with GANs. No need for a norm ball constraint. They look unperturbed to a human observer but break a model trained to resist large perturbations. https://t.co/m8W1WpQQwu pic.twitter.com/CZEEd3Ssjj

— Ian Goodfellow (@goodfellow_ian) May 22, 2018

Beautiful paper from @ukhndlwl performing a battery of experiments to evaluate the long-range dependencies of LSTMs to word order, part of speech, word frequency, and more. Would be awesome to see these tests become part of the standard evaluation of RNNs! https://t.co/MVA8Kug0Fa pic.twitter.com/nMulH3ehlj

— Ari Morcos (@arimorcos) May 22, 2018

"A Spline Theory of Deep Networks"
Expanded version of ICML 2018 paper.

* Links DNs to the theory of vector quantization (VQ) and K-means clustering
* Propose simple cost penalty that leads to significantly improved classification
* ...
https://t.co/EYqWB4n68y #DeepLearning pic.twitter.com/NnsOuqCYAp

— ML Review (@ml_review) May 22, 2018

Re-parameterize all the things! If you want to backprop thru samples from mixture, truncated, Gamma, Beta, Dirichlet, Student-t, or von Mises distributions, this paper has gotchu covered. Also faster than RSVI, another general reparameterization trick based on rejection sampling https://t.co/XZ97q21i6C

— Eric Jang (@ericjang11) May 23, 2018

iVQA: Inverse Visual Question Answering #CVPR2018
By Feng Liu et al.

Generate a question that corresponds to a given image and answer pair.https://t.co/gZi6A2hd05 pic.twitter.com/WHdBgu9V2C

— ML Review (@ml_review) May 23, 2018

Can we state once and for all that any modestly sized combinatorial problem has more configurations than "atoms in the universe" so that we don't have to say it every single time?

— Serkan Cabi (@serkancabi) May 20, 2018

But even here, machine learning itself isn’t the thing. It’s ML + heuristics + data + user experience together with smart product design and a problem that’s worth solving.

Companies which bring all of this together, harmoniously, in important domains, will rule the next decade.

— Jensen Harris (@jensenharris) May 23, 2018

"Theoretical Impediments to Machine Learning With Seven Sparks from the Causal Revolution": Judea Pearl presents seven tasks which he claims are beyond reach of current machine learning systems but have been accomplished using the tools of causal modeling. https://t.co/Z2VBCm0jhn

— hardmaru (@hardmaru) May 23, 2018

I've filled out 1000s of these horrible captchas and not once have they asked to ID pedestrians or cyclists. Only street signs and cars and other detritus of car-dependency. If we are truly doing the work of machine vision for AVs, show me the most vulnerable road users. pic.twitter.com/yl9WsKUPWc

— Jeff Novich 🚴 (@jeffnovich) May 22, 2018

AI cities for the ultra-rich titans of an automated future. Sci-fi short from Import AI this week. Book is on the way! https://t.co/vINoUkULGP pic.twitter.com/LrM8zPaXKo

— Jack Clark (@jackclarkSF) May 23, 2018