TJ Torres

354 posts

TJ Torres

@Teejosaur

Machine learning for work. Pilot, snowboardist, and dog lover.

Oakland, CA Katılım Mayıs 2009

259 Takip Edilen406 Takipçiler

TJ Torres@Teejosaur·12 Tem

I have been spending the past few months developing a genAI project that enables direct prompt —> avatar creation in Roblox. It has been a really fun project to work on and the preview experience released today so I can now share. devforum.roblox.com/t/generate-ava…

English

146

TJ Torres retweetledi

Kevin Patrick Murphy@sirbayes·30 Tem

I am delighted to announce that the camera-ready version of my new book, "Machine Learning: Advanced Topics", is finally available online for free at probml.github.io/book2 (@mitpress will publish the hard copy in 2023.)

English

811

3.7K

TJ Torres retweetledi

Erik Bernhardsson@bernhardsson·16 Ara

This vuln exploit is insane. iMessage pre-renders gifs of any received messages outside a sandbox – trick it to render an obscure Turing complete file format made for printers in the 90s, and trigger a buffer overflow: googleprojectzero.blogspot.com/2021/12/a-deep…

English

TJ Torres@Teejosaur·9 Tem

I really enjoyed this post. It is a perhaps slightly more cynical view than my own on building data teams, but mirrors a lot of my experience and conversations I’ve had with others on the topic. Some great reading from @fulhack

Erik Bernhardsson@bernhardsson

I wrote a blog post about how to build at data team in the form of a story (and it ended up being about 4x longer than I anticipated): erikbern.com/2021/07/07/the…

English

TJ Torres retweetledi

Akshay Agrawal@akshaykagrawal·24 Mar

New paper: Minimum-Distortion Embedding paper: arxiv.org/abs/2103.02559 We introduce a framework (generalizing spectral, PCA, MDS, UMAP ...) that leads to new embeddings code: pymde.org PyMDE lets you fit custom embeddings w/constraints on a GPU via @PyTorch

GIF

English

348

1.7K

TJ Torres@Teejosaur·25 Kas

Deciding to share some exciting career news: As of next Monday I’ll be joining the @Roblox Data Science team leading efforts for Trust and Safety. I’m really looking forward to take on the challenge of keeping the platform safe and enjoyable for all players.

English

TJ Torres@Teejosaur·22 Ağu

@emollick This makes a lot of sense, but worth noting that adopting a strategy of more frequent smaller trials contrasts the way we reward progress in industry right now. PMs are often judged on % successful outcomes, not volume of experimentation. Would require a larger shift in mindset.

English

Ethan Mollick@emollick·22 Ağu

Finding breakthroughs at large companies: A/B tests at Bing show most tweaks have tiny benefits requiring large tests to identify. Instead, run more small & less accurate tests: the top 2% of ideas are responsible for 74.8% of the gains. You can screen for radical ideas quickly!

English

156

788

TJ Torres retweetledi

Google AI@GoogleAI·28 Tem

Check out a new approach to vector similarity search that compresses dataset vectors to enable fast approximate distance computation while significantly boosting the accuracy of database queries compared to prior approaches. Learn about it here: goo.gle/30U126j

English

252

824

TJ Torres retweetledi

Google DeepMind@GoogleDeepMind·16 Haz

Moving away from negative pairs in self-supervised representation learning: our new SotA method, Bootstrap Your Own Latent (BYOL), narrows the gap between self-supervised & supervised methods simply by predicting previous versions of itself. See here: bit.ly/30MCugQ

English

218

764

TJ Torres retweetledi

François Chollet@fchollet·12 Haz

If you are a Black ML developer or data scientist in the US, show me your open-source projects / portfolio, and I will retweet. You deserve attention, respect, and opportunity.

English

231

1.3K

TJ Torres retweetledi

Google DeepMind@GoogleDeepMind·6 May

For students and others interested in expanding their knowledge of AI during this period, we thought it might be helpful to ask our researchers what they consider to be the most impactful and insightful resources available to use #AtHomeWithAI (1/9)

English

930

TJ Torres@Teejosaur·25 Nis

Really nice discussion on the intricacies of training in high cardinality multi-task settings.

Eric Jang@ericjang11

This talk by @karpathy youtu.be/IHH47nZ7FZU has convinced me that Tesla is several years ahead of most CV labs in regards to pushing the limits of DL. Commonplace questions like "how do you do early stopping for a multi-task model?" are non-trivial when at scale.

English

TJ Torres retweetledi

Ankur Handa@ankurhandos·20 Şub

PointRend: Image Segmentation as Rendering arxiv.org/abs/1912.08193 They use adaptive sampling around regions of uncertainty to refine the labels going from coarse to fine resolution inspired by graphics based rendering techniques. code: github.com/facebookresear…

English

333

TJ Torres retweetledi

John McDonnell@johnvmcdonnell·10 Şub

New post, how a monkey can beat the S&P 500… I'm honestly still trying to wrap my head around why cap weighting is so bad, but at least on the historical record it's pretty clear that it's a losing strategy. jvmcdonnell.com/2020/02/03/how…

English

TJ Torres retweetledi

Ned 'no longer here' Potter@ned_potter·16 Oca

Cows make milk. They milk themselves. Other cows check the milk (for free). Cows - get this - PAY THE FARMER to take the milk away. Then the farmer (you won't believe this, honestly) sells the milk *back to the cows.* #academicpublishing

English

103

8.3K

TJ Torres retweetledi

Chris Said@Chris_Said·10 Oca

I just wrote a 3-part blog post on optimizing sample sizes in A/B testing. It's kind of a longread, so if you just want the elevator pitch, here's a tweetstorm (1/20): chris-said.io/2020/01/10/opt…

English

359

TJ Torres retweetledi

Ilya Sutskever@ilyasut·13 Ara

The DotA paper is out!!!

OpenAI@OpenAI

We're releasing "Dota 2 with Large Scale Deep Reinforcement Learning", a scientific paper analyzing our findings from our 3-year Dota project: openai.com/projects/five/ One highlight — we trained a new agent, Rerun, which has a 98% win rate vs the version that beat @OGEsports.

English

362

TJ Torres@Teejosaur·12 Ara

@jxnlco 😂 I think last time for me was in E&M with multipole expansions.

English

jason liu@jxnlco·12 Ara

@Teejosaur I have not heard that word in years.

San Francisco, CA 🇺🇸 English

TJ Torres@Teejosaur·11 Ara

Ok this is a pretty novel idea (at least to me) and seems to have some really nice results in terms of long term dependence.

hardmaru@hardmaru

Found this gem @ #Neurips2019! Using orthogonal property of Legendre polynomials, Legendre Memory Units (LMU) can efficiently handle temporal dependencies spanning 100k timesteps, converge rapidly and use fewer internal state-variables compares to LSTMs. papers.nips.cc/paper/9689-leg…

English

TJ Torres retweetledi

hardmaru@hardmaru·11 Ara

Deep Equilibrium Models Their method is equivalent to running an infinite depth (weight-tied) feedforward net, but has the notable advantage that they can analytically backpropagate through the equilibrium point using implicit differentiation. Cool work! arxiv.org/abs/1909.01377

English

190

Keşfet

@mitpress @fulhack @PyTorch @Roblox @emollick @jxnlco @elonmusk @BarackObama