Jonathan Uesato

Deploying ML in high-stakes situations will require new evaluation techniques beyond static hold-out sets. It's exciting to share out views and recent progress on this problem from our team at DeepMind and so many others in the community.

Yesterday we announced the BASALT competition. Why did we make it? TL;DR: It lets us test our ability to build agents that solve fuzzy tasks where rewards are hard to specify. (1/6) Participate: aicrowd.com/challenges/neu… Paper: arxiv.org/abs/2107.01969 Blog: bair.berkeley.edu/blog/2021/07/0…

Our work on alignment of language agents - how AI designers can misspecify what they want an AI to do with examples in the language setting. w/ @tom4everitt @weidingerlaura @IasonGabriel Vladimir Mikulik @geoffreyirving Blog dpmd.ai/39oMkt6 Paper dpmd.ai/2Pa8k41

I often hear fellow researchers state that "our world is a (PO)MDP". I vehemently disagree: A (PO)MDP is a convenient model. From Bellman's DP book (1957): "It is important to realize that these are very strong assumptions concerning the nature of the system."

When we make assumptions about what features exist in neural networks, they often prove us wrong. It turns out that 4% of CLIPs final neurons (8% on a liberal interpretation) are focused on geography. I certainly wouldn't have guessed that in advance! #region-neurons" target="_blank" rel="nofollow noopener">distill.pub/2021/multimoda…

@SamuelMLSmith A lot of your works are in my go-tos for whenever people ask for examples of theory influencing practice in deep learning! Glad to have another for the list :)

My two takeaways: 1) The most important thing when training deep networks without tricks, is to get the initialization scheme right! 2) Theoretical work can have tangible practical benefits, but this is much more likely when theorists and practitioners collaborate closely. 4/4

Proud to be a part of NFNets, a new ImageNet SOTA: - does not use BatchNorm, LayerNorm, GroupNorm, anyNorm! - 86.5% top-1 w/o extra data - 89.2% top-1 w/ pre-training - 8.7x faster than EffNet-B7 to same test accuracy arxiv.org/abs/2102.06171 code: dpmd.ai/nfnets 1/4

Releasing NFNets: SOTA on ImageNet. Without normalization layers! arxiv.org/abs/2102.06171 Code: dpmd.ai/nfnets This is the third paper in a series that began by studying the benefits of BatchNorm and ended by designing highly performant networks w/o it. A thread: 1/8

@Raza_Habib496 Are there any papers in this space that you like and would be happy sharing?

I continue to be amazed by how little of academic ML research looks how we collect and label data, given that for almost any real application this is the biggest factor in performance.

Come check out verifying previously intractable verification-agnostic networks with SDPs, without relying on any training procedure modifications! Paper: arxiv.org/abs/2010.11645 Code: github.com/deepmind/jax_v… Poster (5-7PM GMT today): neurips.gather.town/app/ws6861DYR4…

We designed REALab, a platform with tampering opportunities integrated into the task dynamics. And developed some cool algorithms to run in it as well!

@srchvrs Yes, precisely

@JonathanUesato Thank you, and by verification you understand the check if we have adversarial examples in a given ball?

Excited to share #NeurIPS2020 papers on efficient and tight neural network verification, based on efficient solvers for LP and SDP relaxations. Implementations of these in JAX are also available as part of the new jax_verify library, described here: bit.ly/2TE1Qcc

@CevherLIONS @DeepMind @BunelR @percyliang @JacobSteinhardt @goodfellow_ian @alexey2004 @pushmeet @sdathath @madeleineudell That's great news! Yes, I'll drop you a line to coordinate. (My email is firstinitiallastname@google.com)

@JonathanUesato @DeepMind @BunelR @percyliang @JacobSteinhardt @goodfellow_ian @alexey2004 @pushmeet @sdathath @madeleineudell We will try to get an implementation done on JAX. It may take a bit of time though. Happy to chat on this separately via email, if it makes sense.

@srchvrs @DeepMind @BunelR @percyliang @JacobSteinhardt @goodfellow_ian @alexey2004 @pushmeet @sdathath Hi, these papers are focused on verifying trained networks, as opposed to training verifiable networks. For the SDP paper, we can verify robustness on CIFAR-10 at {2,4}/255 with small CNNs. Table 1 is probably the fastest place to look if you're interested in details.

@DeepMind @JonathanUesato @BunelR @percyliang @JacobSteinhardt @goodfellow_ian @alexey2004 @pushmeet @sdathath What is exactly the adversarial training, i.e., the size of the epsilon, the number of PGD steps, etc and what's the performance of the trained model on the clean data?

@CevherLIONS @DeepMind @BunelR @percyliang @JacobSteinhardt @goodfellow_ian @alexey2004 @pushmeet @sdathath @madeleineudell It's tricky to integrate with the Matlab version, since our problems are only defined implicitly via autodiff, but if there's a JAX version of SketchyCGAL, we could naturally extend jax_verify to define the affine primitive operations.

@CevherLIONS @DeepMind @BunelR @percyliang @JacobSteinhardt @goodfellow_ian @alexey2004 @pushmeet @sdathath @madeleineudell Hi, we've only tried the straightforward dual ascent approach, but it'd be really interesting to try some of the ideas from those papers, e.g. the smoothness properties inherited by SketchyCGAL could be very beneficial.

Have you ever wondered how to reach 65% robust accuracy on CIFAR-10 (against l-infinity norm-bounded adversaries of size 8/255)? In collaboration with Chongli Qin, @JonathanUesato, @realKingTim and @pushmeet, we are releasing our recipe at arxiv.org/abs/2010.03593. (1/4)

@RichardMCNgo But there's so many things to try. Things rarely work first try in DL - to make them work, you need conviction to keep trying. It would have happened eventually, but it would have taken much longer to discovery + widespread adoption

@JonathanUesato Sure, maybe you individually wouldn't have invented it. But somebody would have tried it, seen that it worked, and spread it around. You don't need maths to imagine a ball rolling down a hill.

I'm usually pretty skeptical about the usefulness of formal proofs in AI and ML. But I'm open to changing my mind. What are the most important proofs in the history of AI? In particular, I'm interested in cases where we couldn't have achieved good empirical results without them.

@RichardMCNgo I agree with the prediction - the goal isn't plug-in bounds on sample complexity, it's to change the way we think about the algorithm. - Theory shows momentum is a great idea - Theory limitations prevent analyzing all scenarios - Still evidence we should still use it in practice

@JonathanUesato My guess (without having investigated) is that such proofs rely on assumptions which aren't very important for empirical success. E.g. concrete prediction: you could get similar performance with learning rate schedules that provably converge, and ones which don't.

@RichardMCNgo I think the clean split between quadratic and linear convergence for momentum provided a lot of the early excitement for it. E.g. I don't think you get as much early exploration into methods like Momentum and Adam if those results don't exist

@RichardMCNgo I don't think it's prima facie obvious momentum is a good idea. E.g. if it hadn't been invented, I don't think I'd have discovered it over the course of NN training. Probably would have to look at simpler models, like optimizing a quadratic.