Day 30: Tap of Occasional Insight

Congrats @jnhwkim and team on winning the VQA challenge at #CVPR2018.
Read their paper "Bilinear Attention Networks" at https://t.co/sqXNysYxcv
PyTorch based code at: https://t.co/XIOwVHfLtO https://t.co/FGOR3Pe9Rl

— PyTorch (@PyTorch) June 19, 2018

@PyTorch geometric by @rusty1s et al.https://t.co/rhyvOyjppS pic.twitter.com/mOWWuFClUD

— Brandon Amos (@brandondamos) June 18, 2018

Deep-learning-free text embeddings. Surprisingly simple text embeddings suffice to match the performance of much more sophisticated methods for capturing the meaning of text. https://t.co/V3ygzbCthX

— Sanjeev Arora (@prfsanjeevarora) June 18, 2018

Lovely animation of hierarchical clustering by @r2d3us! https://t.co/l1tiYen3fX pic.twitter.com/qPhsS6vdMQ

— Mike Bostock (@mbostock) June 18, 2018

Power analysis and NIH-style statistical practice: What’s the implicit model? https://t.co/loQSW2BMaW

— Andrew Gelman (@StatModeling) June 18, 2018

New approximate nearest neighbor benchmarks https://t.co/VlEMJuEo6I (blog post)

— Erik Bernhardsson (@fulhack) June 18, 2018

I could not be more thrilled to share our paper "Using Search Queries to Understand Health Information Needs in Africa", joint work with Shawndra Hill, Jenn Wortman Vaughan, Peter Small & Andy Schwartz.https://t.co/R4PxcbCgLb

1/n pic.twitter.com/Lfeg1iKx7Y

— Rediet Abebe (@red_abebe) June 19, 2018

Evaluating ELMO embeddings: Meaningful gains on many tasks. "We also show, however, that we are still far away from a universal encoder that can perform consistently across several downstream tasks." by Perone et al 2018 https://t.co/fK93KiDgIX …

— Leonid Boytsov (@srchvrs) June 19, 2018

Just released "CoNaLa", a dataset/contest for broad-coverage generation of programs from English commands: https://t.co/uiDSg3zWIs
2,879 manually annotated examples, and 600k mined from StackOverflow to increase coverage; super-excited to bring NL->Code to the open domain! pic.twitter.com/ah1U6Ld2lL

— Graham Neubig (@gneubig) June 18, 2018

I watched an IBM AI debate two humans in real time and it held its own. I was impressed — and maybe even a little unsettled — by its rhetorical tactics. https://t.co/HGsN7vJpze pic.twitter.com/dWEuVdbggH

— Dieter Bohn (@backlon) June 19, 2018

IBM has decided that Project Debater is a “she.” I decided to go with the “It” pronoun. I spent way too long internally arguing with myself about who gets to decide the appropriate gender pronouns for AI personas. I thought I was 100% “It’ no matter what but now I am conflicted?

— Dieter Bohn (@backlon) June 19, 2018

When research concerned with the potential privacy impact on millions (eventually billions) of users is disregarded, that should give us all pause. The research we ignore in machine learning today will likely make headlines in 2026, impacting billions as well. https://t.co/5UllkUinII

— Smerity (@Smerity) June 18, 2018

The research we will regret ignoring isn't going to be a given branch of neural architectures - it'll be how the models we unleash on the world are destined to impact fairness, accountability, and transparency far beyond what we ever initially envisioned.

— Smerity (@Smerity) June 18, 2018

Just released our #ICML2018 paper "Gated Path Planning Networks" as well as a @PyTorch implementation for replicating our experiments:https://t.co/ZMElrvKLufhttps://t.co/qFGAEEfIyG
- with Emilio Parisotto, @dchaplot, Eric Xing, @rsalakhu

See you @icmlconf in Stockholm! pic.twitter.com/IwPfMH15p2

— Lisa Lee (@rl_agent) June 19, 2018

Zero-shot Recognition via Semantic Embeddings and Knowledge Graphs (#CVPR 2018)

Githubhttps://t.co/vuXAN3JpYv
Arxivhttps://t.co/SxbsaBLHRf pic.twitter.com/VuBgnBupNi

— ML Review (@ml_review) June 18, 2018

"A More General Robust Loss Function"

Two-parameter loss function that generalizes many existing one-parameter robust loss functions: the Cauchy/Lorentzian, Geman-McClure, Welsch, and generalized Charbonnier, pseudo-Huber, L2, and L1 loss functionshttps://t.co/rVV3QN3PJJ pic.twitter.com/txEbJQMJiV

— ML Review (@ml_review) June 19, 2018

Best student paper award for project that began as a CS224U project. https://t.co/4JB4BQ5J5I

— Stanford NLP Group (@stanfordnlp) June 18, 2018

Our new paper from Araya by @ildefons and myself, now on arXiv.
"A unified strategy for implementing curiosity and empowerment driven reinforcement learning"https://t.co/cugFP9kza8

— Ryota Kanai (@kanair) June 19, 2018

An attempt to reproduce WaveNet-based text-to-speech synthesis https://t.co/50xRIcChFh #mlm_kansai 先週のイベントでのLT資料公開しました。ウェーブネットに関して簡単にお話してきました

— 山本りゅういち / Ryuichi Yamamoto (@r9y9) June 18, 2018

Read the paper Universal Language Model Fine-Tuning for Text Classification by @jeremyphoward and @seb_ruder. Exciting times in the NLP world. Have written a blog post explaining the working of the ULMFiT method. You can find it here - https://t.co/45nxVW6dJD

— Yashu Seth (@yashuseth) June 16, 2018

The long-awaited "Visual Introduction to #MachineLearning Part II" is now out, focusing on model tuning and the bias-variance tradeoff. #datavizhttps://t.co/5uhkaHVaUl pic.twitter.com/YocmhXd6ZB

— Randy Olson (@randal_olson) June 18, 2018

bounceR 0.1.2: Automated Feature Selection https://t.co/BgNPjaUwFB #rstats #DataScience

— R-bloggers (@Rbloggers) June 18, 2018

JupyterHub 0.9 (multi-user notebook server) released: https://t.co/ufFBUzEQ3n

— Min RK (@minrk) June 18, 2018

In #Python3, the print function has options
* 'sep' to separate the arguments
* 'end' to be printed after the last argumenthttps://t.co/Jm8LgiGnAJ pic.twitter.com/V56ARBc7ku

— Daily Python Tip (@python_tip) June 18, 2018

10 common security gotchas in #Python and how to avoid them. #programminghttps://t.co/mDxC583eu5 pic.twitter.com/U084t3D3TB

— Randy Olson (@randal_olson) June 18, 2018

ICYMI, 👍 overview:
"Enterprise Dashboards with R Markdown" by @nwstephens
https://t.co/YwE2q5K8dE #rstats #rmarkdown #dataviz pic.twitter.com/PvLn7vA22V

— Mara Averick (@dataandme) June 18, 2018

Celebrating 2 amazing self-taught AI talents@johnolafenwa @OlafenwaMoses
Check their work https://t.co/kuuO7h1nKn
TorchFusion, a modern @PyTorch based deep learning framework to accelerate research& development of AI systems.
They will be at the AIHub to share their work 29/05 pic.twitter.com/qtBDJQn4OZ

— DataScienceNigeria (@DataScienceNIG) June 18, 2018

Watson Health is scaling back, Facebook M never delivered, driverless cars are facing challenges.

Not saying another AI winter is coming, but overpromising is certainly something that could lead to one. https://t.co/6opUPFM4OQ

— Gary Marcus (@GaryMarcus) June 19, 2018

Epistemological variations pic.twitter.com/C9xQqAXEuv

— Philosophy Matters (@PhilosophyMttrs) June 17, 2018

Good documentation makes the reader feel smart. Great documentation makes the reader feel like they just put on the One Ring. https://t.co/Tqgaue1YEJ

— Brandon Rohrer (@_brohrer_) June 18, 2018

This sounds very encouraging. Designed to take people who already know how to train DL models and show them how to create end to end solutions. They mention @fastdotai as a good foundation for this. https://t.co/TCVMz7GCKU

— Jeremy Howard (@jeremyphoward) June 18, 2018

people who insist that y axes must go down to zero aren't allowed to seek medical treatment for a high fever; after all, it's a barely recognizable percentage change in body temperature

— Matthew Zeitlin (@MattZeitlin) June 18, 2018

"Clever machines", as The Economist calls them, will not replace radiologists, any more than auto ated teller machines replaced bank tellers. But it will certainly transform their job for the... https://t.co/KAA2nONjWM

— Yann LeCun (@ylecun) June 18, 2018

Oh wow, these deep learning libraries are so accurate these days pic.twitter.com/c6Xg2ACtNO

— Vicki Boykis (@vboykis) June 19, 2018

We’ve developed a technique for protecting voice interfaces by identifying whether or not the source of the audio was a human being or a speaker (e.g., from your TV, smartphone, #iot device, etc). 1/ @uf_fics @ufcise pic.twitter.com/MtlDDm2kw0

— Patrick Traynor (@patrickgtraynor) June 18, 2018

Google Cloud Platform slashed GPU pricing this month. P100s for less than 50c an hour, K80s for pennies... 🔥 https://t.co/C7MvhZdiSt pic.twitter.com/rKDJl5di9h

— hardmaru (@hardmaru) June 18, 2018

I do almost all of my research on GCP Ubuntu virtual machines these days using the open source stack. The World Models project was done entirely using a P100 instance, and a 64-core CPU instance on a Google Cloud VM: https://t.co/UmCE5L6vbH

— hardmaru (@hardmaru) June 18, 2018