Pinned Tweet
Interested in understanding the role regularization plays for (distributional) robustness?
Check out my paper “Distributional Robustness with IPMs and links to Regularization and GANs” (accepted to NeurIPS2020) with general IPM results. arxiv.org/pdf/2006.04349.pdf #NeurIPS2020
Hisham Husain retweeted
📢Pre-print Alert📢
There is a new algorithm for projection-free (Online) Convex Optimization. No need for expensive Euclidean projections whenever the iterates step outside the feasible set. It achieves the optimal regret bound unlike Frank-Wolfe arxiv.org/pdf/2111.05818.pdf
Hisham Husain retweeted
In Meta-RL, the *agent* is responsible for collecting the data that we meta-train on. If it doesn't do this well, meta-training can fail! 😱
We address this meta-exploration problem in our #ICML2021 paper. Come by our poster this Friday 5-7AM BST!
ALT Screenshot of the title slide of our ICML presentation. At the top the title of the paper in large text: "Exploration in Approximate Hyper-State Space for Meta-RL". At the bottom are the names and picture portraits of the authors: Luisa Zintgraf, Leo Feng, Cong Lu, Maximilian Igl, Kristian Hartikainen, Katja Hofmann, Shimon Whiteson. The top right shows the logos of the author's affiliations: University of Oxford, WhiRL lab, and Microsoft Research. A star-shaped field in the middle of the slide highlights the words "A way to solve the meta-exploration problem"!
Hisham Husain retweeted
ICML recording done 🎉 Always fun to build a little recording area 🎥😬🪴
Finally getting better at recording - The entire thing only took me 2 days this time, not 2 weeks 🙌🏼😎
ALT A screenshot of the Slideslive recording interface: me on the left, the first slide of my presentation on the right, and a “start recording” button at the bottom.
ALT A picture of my recording set-up: my propped up laptop on an empty box, plants in the background, and a light strip on the side.
Hisham Husain retweeted
Congratulations to @MengyanZg on your first accepted paper! To be presented soon at #ICML2021 , see arxiv.org/abs/2010.11568.
Hisham Husain retweeted
Half of the heroes in the MCU seem to have PhDs that they're not using. The whole franchise is a story about the academic employment crisis.
Hisham Husain retweeted
Looking forward to present our poster on “Deep Interactive Bayesian RL via Meta-Learning” at #AAMAS in 1 hour! Take a sneak peek below 🤖🐣🤓
Joint work with @smdvln @MLciosek @shimon8282 @katjahofmann. Thinking back to the times of my internship at @MSFTResearch Cambridge! ♥️
Hisham Husain retweeted
Very excited by our recent paper on "Robust Generalised Bayesian Inference for Intractable Likelihoods" with @TakuoMatsubara, @LauchLab and Chris Oates; you can find it here: arxiv.org/abs/2104.07359. [1/10]
Hisham Husain retweeted
Depressing to see my work on AdaGrad constantly missing in the discussion of prior work. I was the first one to prove convergence and adaptivity in stochastic convex and non-convex arxiv.org/pdf/1805.08114v1.p…
I bet if I were a big shot people would not forget my work so easily😞
We're introducing an optimizer for deep learning, MADGRAD. This method matches or exceeds the performance of the Adam optimizer across a varied set of realistic large-scale deep learning training problems. github.com/facebookresearch/…
Hisham Husain retweeted
PMLR Volume 130
Proceedings of AISTATS 2021
Is now available
proceedings.mlr.press/v130/
Hisham Husain retweeted
This is \Huge: Giles Gardam just announced a counterexample to the Kaplansky unit conjecture! The group ring ℤ₂G of the Promislow group has non-trivial units. Congratulations Giles!
Very delighted to share this pre-print: joint work with @mhammediz on risk-monotoncinity (a property often taken for granted) with very interesting findings!
Do you expect the risk of your algorithm to monotonically decrease when trained on more data? Surprisingly, this is not always the case even for ERM. With @syimplectic, we derived general risk-monotonic algorithms arxiv.org/abs/2011.14126, resolving parts of a COLT19 open problem.
Hisham Husain retweeted
The mathematician Karl Weierstrass was born #OTD, Halloween, in 1815. The fools at the academy all said he was mad, but in 1872 he announced that he had succeeded in creating a ~monster~.
Image: Smithsonian Institution Libraries
Hisham Husain retweeted
4 days left. Just in time for the quality ML theory papers that lost the #NeurIPS2020 lottery.
PSA: ALT deadline is 2 weeks away and the submission server is open for business. algorithmiclearningtheory.or…
Hisham Husain retweeted
Highly recommended: draw your figures with the excellent free Ipe sofware ipe.otfried.org/ and export into pdf for inclusion in latex files. Can be used for slides too... RT!
Hisham Husain retweeted
The virtual website where to watch papers and keynotes videos is now available to all participants!
Subscriptions are still open at aistats.org. #AISTATS2020
Hisham Husain retweeted
1/3
Why is it so hard to design Continual Learning algorithms that do not forget catastrophically?
We give the answer: Such algorithms would solve an NP-hard problem with perfect memory (arxiv.org/abs/2006.05188)
ICML paper with @syimplectic @tommy_da_cat
