Hoo Chang Shin

The journey is never complete.

[Hoo Chang Shin @hugeiezzy]

Can a (very) large language model like GPT-3 (just magically) "do it all"?
I guess not.

In our recent BioMegatron paper we show it's probably better and more important to pre-train on your domain text, over model/text-corpus size:
https://t.co/sl1Rem9Qir

#emnlp2020

[hardmaru @hardmaru]

Neural Processes in @PyTorch: Short blog post explaining “Neural Processes” (https://t.co/qXF9MuXgsx), connection to VAEs, and highlights shortcomings with the approach (with suggestions to make it work better). https://t.co/BhzytA9IPh

[Snowkylin @snowkylin]

An online TensorFlow handbook (https://t.co/eiWuQNCNPS) based on Eager Execution to help more developers to get started with TensorFlow as painlessly as possible. Both Chinese and English version are available online.
Open-sourced on GitHub: https://t.co/Op3vMMIfkr

[Taiki SEKII @TaikiSEKII]

YOLO meets OpenPose: ~200 FPS, multi-person pose estimation.
My ECCV paper is now online.
- Paper: https://t.co/qc7DD7xb35
- Poster: https://t.co/pQbulR58o2
- YouTube: https://t.co/Em2Lmc1Fqq
#ECCV2018

[Yaroslav Bulatov @yaroslavvb]

The code to reproduce "Imagenet in 18 minutes" is posted https://t.co/uuPAsSb4id , please let me know if you run into any problems. It actually becomes "Imagenet in 12 minutes" if we use 74.9% top1 used in Chainer's "Imagenet in 15" paper, last few bits are the hardest.

[TensorFlow @TensorFlow]

Check out this helpful walkthrough of Multi-GPU training with Custom Estimators by @krasul .

Read it here ↓ https://t.co/SvVZbO3xbB

[TensorFlow @TensorFlow]

One hot encoding may look inefficient at first, but it's actually very effective for training. In this week's tip, @lmoroney dives into the code to show why this is and how the argMax function can help you find desired values.

Watch it here → https://t.co/BU3t6t28EU

[Carles Gelada @carlesgelada]

Check out our new RL framework! Doing RL research used to be hard, and that won't change. But with a solid codebase you can focus on solving the problems that really matter.
The core team is @marcgbellemare, @pcastr, Deep Moitra and I. https://t.co/PvkP9Pg5Ex

[Gabriel Peyré @gabrielpeyre]

Representation theory defines a Fourier transform on groups. Finite commutative groups correspond to the classical discrete Fourier transform. https://t.co/Kany2ueLmN https://t.co/agPR4YMypg

[Gabriel Peyré @gabrielpeyre]

The Fisher metric is the unique Riemannian structure on parametric families of densities which is invariant/covariant under reparametrization of the samples/parameter spaces.
For 1D Gaussians, it corresponds to the Poincaré hyperbolic half plane. https://t.co/paVsLKPZmL

[Kyunghyun Cho @kchonyc]

very interesting! https://t.co/jRXM7cpw1Z

[PyTorch @PyTorch]

New seq2seq architecture - jointly encodes source and targets into a 2D ConvNet. No enc/dec or explicit attention.
Outperforming ConvS2S and Transformers on IWSLT'14 de<->en, with 3 to 8 times less parameters
from @melbayad and team
https://t.co/8NmiwmnhI2
https://t.co/LqUYynj8vB

[Soumith Chintala @soumithchintala]

"Everybody Dance Now" from Caroline Chan, Alyosha Efros and team transfers dance moves from one subject to another
The only way I'll ever dance well. Amazing work!!!
https://t.co/qFh5G4JjDC
https://t.co/8QNcbNwHZb

[Durk Kingma @dpkingma]

Decent looking reproduction in Chainer of our face synthesis results with the Glow model: https://t.co/YDEUW5U18F. I don't read Japanese but according to Google Translate the author's blog (https://t.co/JgZpTaqw9K) is appropriately titled "Defense technique against dark magic" :)

[TensorFlow @TensorFlow]

Check out TensorFlow Probability's new home page. Complete with @GoogleColab notebook examples (e.g., How to build hierarchical Bayesian mixture models) and API docs.

See for yourself ↓ https://t.co/PLVt3azv3r

[Andrej Karpathy @karpathy]

Discovering (paradoxically late in life) that I get more out of textbooks than books and that I don’t have to stop buying them just because I’m out of school. Good reading list pointers: https://t.co/ZDaqr3YQbW

[Eric Jang @ericjang11]

Happy birthday to this fantastic overview of GAN theory by @lilianweng

https://t.co/NqnkUvPggF

Also check out her other blog posts, they are a very helpful resource for practitioners!

[Jihyeon Kim (김지현) @simnalamburt]

와 gdb-dashboard 쩐다 그 괴상한 gdb 내장 TUI보다 훨씬 낫다
https://t.co/gaRQcbKSHr

지금까지 거의 모든 코딩을 터미널에서 다 해오곤 했지만 gdb 딱 하나만 너무 그지같아서 GUI에서 해왔었는데, 이런거 도움 잘 받으면 터미널에서도 잘 할수 있을듯 싶다

[Google AI @GoogleAI]

Check out Universal Transformers, new research from the Google Brain team & @DeepMindAI that extends last year's Transformer (a neural network architecture based on a self-attention mechanism) to be computationally universal. https://t.co/7loIFE9msM