🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Min-max optimization (used among other applications in GANs, and adversarial training more broadly) is empirically challenging. We show why min-max optimization is hard in the following paper with Stratis Skoulakis and Manolis Zampetakis: arxiv.org/abs/2009.09623

Thanks Gautam!!

Samples from high-dimensional distributions can be scarce or expensive to acquire. Can we meaningfully learn them from *one* sample?!? In new work w/ Yuval Dagan, Nishanth Dikkala, &Vardis Kandiros, we show how to learn Ising models given a single sample arxiv.org/abs/2004.09370

Have you wondered how to win the game "Twenty questions" with lies? youtu.be/FxKKVvtyLDw Now on ITCS, joint with Yuval Filmus Yuval Filmus, Daniel Kane and Shay Moran

The following paper (accepted to STOC 21) studies sampling algorithms on a stream, against an adversary that has access to the internal state of the algorithm. Further, it resolves an open question in online classification. arxiv.org/pdf/2101.09054…

A statistics postdoc has solved an important "slicing" problem in convex geometry. Erica Klarreich reports in her terrific new piece for Quanta Magazine. quantamagazine.org/statistics-pos…

The ACM SIGecom Dissertation Award for 2020 goes to Manolis Zampetakis for his thesis "Statistics in High Dimensions without IID Samples: Truncated Statistics and Minimax Optimization" advised by Constantinos Daskalakis at Massachusetts Institute of Technology (MIT). Congratulations! More details: sigecom.org/award-phd.html

We've analyzed differential-privacy mechanisms that add bounded noise, and it usually outperforms the Gaussian mechanism! (if you consider a uniform bound on the noises of > 1000 queries) new revision: arxiv.org/pdf/2012.03817… (Joint with Gil Kur)

Happy to share with you a new paper with Steve Hanneke and Yishay Mansour: A Characterization of Semi-Supervised Adversarially-Robust PAC Learnability. arxiv.org/abs/2202.05420

Wow! Brukhim, Carmon, Dinur, Moran, and Yehudayoff just resolved a long-standing open question on multi-class learning. They show that learnability is equivalent to the (non)-existence of "pseudo-cubes" over large enough samples (DS-dimension) (1/3)

DALLE-2 has a secret language. "Apoploe vesrreaitais" means birds. "Contarra ccetnxniams luryca tanniounons" means bugs or pests. The prompt: "Apoploe vesrreaitais eating Contarra ccetnxniams luryca tanniounons" gives images of birds eating bugs. A thread (1/n)🧵

New ICML paper: Score-Guided Intermediate Layer Optimization (SGILO). We train diffusion models on the latent space of StyleGAN and we show provable mixing of Langevin Dynamics for random generators. Reconstructions for *extremely sparse* (<1%) measurements. A thread🧵(1/N)

Announcing Soft Diffusion: A framework to correctly schedule, learn and sample from general diffusion processes. State-of-the-art results on CelebA, outperforms DDPMs and vanilla score-based models. A 🧵to learn about Soft Score Matching, Momentum Sampling and the role of noise

We studied the setting of semi-supervised adversarially robust PAC learning. Perhaps surprisingly, we show that the labeled sample complexity can be arbitrarily smaller than the unlabeled one, and controlled by a different complexity measure. #NeurIPS2022 arxiv.org/abs/2202.05420

At COLT 2023! Learning and Testing Latent-Tree Ising Models Efficiently. Given i.i.d samples from the leaves of some evolutionary tree, we learn it in TV. Proof relies on a tensorization argument for latent models. w. V Kandiros, C Daskalakis and D Choo. arxiv.org/abs/2211.13291

Stable Diffusion and other text-to-image models sometimes blatantly copy from their training images. We introduce Ambient Diffusion, a framework to train/finetune diffusion models given only *corrupted* images as input. This reduces the memorization of the training set. A 🧵

Exciting work w/ Yuval Dagan Maxwell Fishelson Noah Golowich on efficient algos for no-swap regret learning and, relatedly, correlated eq when the #actions is exponentially large/infinite. While classical works point in the opposite direction, we show that this is actually possible!