Utkarsh Singhal

Pi models are now running in production settings, in collab with @Ultraroboticsco and @weaverobotics. We see: - much higher autonomy with pi-0.6 over using pi-0.5 - fewer mistakes & higher throughput from incorporating data in pre-training Blog post: pi.website/blog/partner

It was absolutely disgraceful seeing the venue people just rudely cut the mic. Even when @breadli428 repeatedly beg for just 3 more minutes and he will be done, the person simply repeated robotically: you should have been done. I was clearing rooms one by one. It’s the event organizer’s fault. Absolutely no respect at all.

@antoniloq Congrats! :D

🤖🇮🇹🥳

@_onionesque Also the finding in question is explained in the first ~5-ish pages of the report. They didn’t even read the summary!

Very misleading post. Low ROI from generative AI pilots is mostly a reflection of immature AI strategy (still very rudimentary in most companies across verticals) and execution (better, but still lagging), not because of the state of generative models.

@SwayStar123 Matches my observations too! One thing I've found interesting is that landscapes/scenes converge the fastest. Is it a coincidence that the first large-ish generative models (e.g., iGANs) were for landscapes?

Tried adding contrastive flow matching loss. Worse FID @ 100k. I think whats happening is that CFM might only be beneficial for low step sampling? I am doing these evaluations at 250 steps whereas CFM mysteriously only reports at like 50 steps or something. You can also see the FID gain tapers out in one of their chart

Happening now 😀 ICML poster session, E. Hall A-B (E-2203)

Check out our work @ ICML tomorrow! I can't join in person 🥲, but I'm genuinely excited to share this work. FoCal hits the core issues I've encountered working on invariance over the last few years: complex transforms, scaling to foundation models & data-driven invariance 😀

@aaron_defazio @torchcompiled Would love to read the paper! I have seen similar phenomena with noisy gradient estimates. My guess before seeing the paper: noise accumulates in the weights over many batches until it starts producing 2nd order (hessian) effects.

@torchcompiled I have been working on a NeurIPS paper that fully explains this phenomenon. The fix is one line of code! I’ll share it as soon as I’m able.

why does grad norm continually grow with training? this is mad unintuitive. absolutely destroying my mental model of 2D convex bowl optimization.

@AllanZhou17 my mute list keeps expanding 🥲

@EliSennesh Is this even a hot take? Since all EBMs are probabilistic models and vice versa? (The only difference being implementation, which i will conveniently ignore :) )

Hot take: "The cortex is an energy-based model" is a reasonable generalization of "The cortex is a Bayesian generative model".

Our gloriously picked cherries. Their wicked failure cases.

Hi all, I'll be at NeurIPS from Tuesday to Sunday. I would love to chat with people about invariance, test-time optimization, and vision. Please DM me if you'd like to talk (or catch up)

@JamesAllingham @kayembruno @shreyaspadhy @JaviAC7 @DavidSKrueger @eric_nalisnick @jmhernandez233 Wonderful work! I love the beautifully drawn graphics especially. :) I've been working on a similar line of research (utkarsh.ai/projects/learn…). There is so much exciting stuff to be done. It would be great to chat at NeurIPS!

A big shoutout to all of my amazing collaborators who made this paper happen! @kayembruno @shreyaspadhy @JaviAC7 @DavidSKrueger Richard Turner @eric_nalisnick @jmhernandez233 This project was 2 years from conception to publication, but I am very excited about the result!

I'll be at NeurIPS next week, presenting our work "A Generative Model of Symmetry Transformations." In it, we propose a symmetry-aware generative model that discovers which (approximate) symmetries are present in a dataset, and can be leveraged to improve data efficiency. 🧵⬇️

@LiyuanLucas Would love to chat too!

Join Microsoft Research's Deep Learning team in Redmond as a Summer 2025 intern! 🎓 Apply at shorturl.at/wlQot 📍 I'll be at #NeurIPS2024 next week - let's connect and chat! Please help us share this post in your networks : ) #DeepLearning #Internship #MSR

@miniapeur I was told there would be cookies

Tell me your story. How did you get into research?

@wgussml Really cool observations! Might be related to some of @thisismyhat 's recent work :) Personally I wonder if this is because (1) the loss functional is convex in function space; (2) averaging gradients ≈ averaging learned functions, at least for MNIST.

in this experiment we take n random initializations and train them on disparate data but average their gradients and update them all with this single mean gradient (this is not an ensemble)! surprisingly on MNIST, this mean gradient leads to faster convergence (even when adjusting for number of examples seen) !???? this property doesn’t hold in general and on cifar eg the opposite is true (which you should expect) it’s interesting to me that the loss landscape of mnist is such that the “mean path” of adam yields a global optima. likely a statement of dataset simplicity but surprising nonetheless

@MaxAifer Sounds exicting!

I'm launching a new reading group on spaces this weekend to talk about excerpts from @sirbayes's book on probalistic machine learning probml.github.io/pml-book/book2…. Who wants to join?