Andrew Gibiansky

Working on neural speech synthesis at Facebook. Formerly founder @ Voicery, research scientist @ Baidu.

[Peyman Milanfar @docmilanfar]

Our eyes are a lot more optically complex than we think

Mitochondrial bundles *inside the photoreceptors* seem to work as tiny lenses that help focus light at the local scale -as if mammalian eyes are also compound eyes like insects and other arthropods

https://t.co/GfyxzH7nvh

[Safe Street Rebel @SafeStreetRebel]

Tonight we protested the city’s complete mismanagement of Valencia Street. While cars were loading in the bike lane, we paused car traffic to create a safe route for bikes.

[Rafael Mandelman @RafaelMandelman]

We love Bernie and Bernie's coffee - a true Noe Valley community institution! https://t.co/nVf4hVukdv

[Andrew Gibiansky @agibiansky]

@RafaelMandelman I live in your district. Please do better. Or at least, don't prevent the other supervisors from doing better in their districts.😠 CEQA is not a tool for concerns about gentrification or vague complaints about community input; stop using it as such.

[Matt Haney @MattHaneySF]

D8 is averaging less than 3 new affordable units permitted or entitled per year over the last 10 years. Less than 3.

[heiner @HeinrichKuttler]

Python _can_ be made GIL-free. https://t.co/LPrHsmfwzk 🤯

[Bryan Catanzaro @ctnzr]

Language models do just keep scaling!

Today we’re announcing Megatron-Turing NLG: a 530B parameter language model.

Joint work from @NVIDIA and @Microsoft. Trained using Megatron and DeepSpeed on DGX SuperPod.

https://t.co/xFN7c8xYQR

[AK @_akhaliq]

Video Autoencoder: self-supervised disentanglement
of static 3D structure and motion
pdf: https://t.co/YhVSZvUZu2
abs: https://t.co/lunP7FmxvH

[Tim Dettmers @Tim_Dettmers]

I am excited to share my latest work: 8-bit optimizers – a replacement for regular optimizers. Faster 🚀, 75% less memory 🪶, same performance📈, no hyperparam tuning needed 🔢. 🧵/n

Paper: https://t.co/V5tjOmaWvD
Library: https://t.co/JAvUk9hrmM
Video: https://t.co/TWCNpCtCap

[thelousylinguist @lousylinguist]

Phrasal verbs, amiright?

Q: Can you kick that ball?
A: I can kick the shit outta that ball.

Q: Can you pick up that ball?
A: *I can pick the shit up outta that ball.
A': *I can pick up the shit outta that ball.

Are my intuitions shared?

[Andrew Gibiansky @agibiansky]

I saw "Variational Inference at Glacial Scale" on Arxiv, and thought it was a clever quip about some incredibly slow generative model... Nope. "Posterior over ice softness parameters of an ice sheet model..." Literally glacial.

[AK @_akhaliq]

NeuralMVS: Bridging Multi-View Stereo and Novel
View Synthesis
pdf: https://t.co/PboI1vZ3HB
abs: https://t.co/D1iKTb7CeJ

proposed a network that jointly resolves scene geometry and novel view synthesis from multiview datasets and is supervised only by image reconstruction loss

[Drew Jaegle @drew_jaegle]

Perceiver IO: a new & improved Perceiver for all of your input and output needs.

Check out the paper (https://t.co/7rXffHqp8f) & code (https://t.co/7AZigAlTeh) for more on masked & multi-task language modeling, optical flow, ImageNet, StarCraft, video+audio+label autoencoding!

[arXiv Sound @ArxivSound]

``The History of Speech Recognition to the Year 2030. (arXiv:2108.00084v1 [https://t.co/JosjykNw75]),'' Awni Hannun, https://t.co/LkmElslFEn

[Yunzhu Li @YunzhuLiYZ]

Introducing “3D Neural Scene Representations for Visuomotor Control”!

https://t.co/IxBEmNsGxO (w/ video!)

We combine implicit neural scene representations with intuitive physics models, enabling visuomotor control of dynamic 3D scenes from out-of-distribution viewpoints. (1/7)

[Andrew Gibiansky @agibiansky]

@ctnzr That would be great! Wow, that is such a perfect name for him. Very clever kids :) I love pun pet names. A friend has a dog named Bella (from German "bellen", to bark) and a wine-loving friend has a cat named Mos-gato (Spanish "gato"). Clef the treblemaker is great!

[Andrew Gibiansky @agibiansky]

Not quite sure what's going on, but the night of medical insomnia doesn't sound fun. Definitely feeling the rap lyric angst though.

[Andrew Gibiansky @agibiansky]

Fun paper: DeepRapper: Neural Rap Generation with Rhyme and Rhythm Modeling. As you can guess, modeling is not the interesting part of rap lyric generation – you use an autoregressive transformer LM. https://t.co/mOHci2I8iG

[Andrew Gibiansky @agibiansky]

Fun paper to read. Unfortunately, my Mandarin vocabulary of ~200 words or so is not enough to appreciate the quality of the generated rap. The English translations don’t help either. https://t.co/Ed88i0idHu