לוח שנה משולב

Modern machine learning systems operate in regimes that challenge classical learning-theoretic assumptions. Models are highly overparameterized, trained with simple optimization algorithms, and rely critically on how data is collected and curated. Understanding the limits of learning in these settings requires revisiting both the computational and statistical foundations of learning theory.

 

A central question in learning theory asks which functions are tractably learnable. Classical complexity results suggest strong computational barriers, motivating a focus on “learnable subclasses” defined by properties of the target function. In this talk, I argue for a different perspective by emphasizing the role of the training distribution. Fixing the learning algorithm (e.g. stochastic gradient descent applied to neural networks), I show that allowing a “positive distribution shift”, where training data is drawn from a carefully chosen auxiliary distribution while evaluation remains on the target distribution, can render several classically hard learning problems tractable.

 

Beyond computational considerations, I then study statistical limits of learning in modern, overparameterized models using stochastic convex optimization as a theoretical framework. While classical theory often suggests that successful generalization requires avoiding memorization, I show that memorization is in fact unavoidable: achieving high accuracy requires retaining nontrivial information about the training data and can even enable the identification of individual training examples. These results reveal fundamental privacy–accuracy tradeoffs inherent to accurate learning.

 

Bio:

 

Idan Attias is a postdoctoral researcher at the Institute for Data, Econometrics, Algorithms, and Learning (IDEAL), working with Lev Reyzin (University of Illinois Chicago), Nati Srebro, and Avrim Blum (Toyota Technological Institute at Chicago). He obtained his Ph.D. in Computer Science under the supervision of Aryeh Kontorovich (Ben-Gurion University) and Yishay Mansour (Tel Aviv University and Google Research).

 

His research focuses on the foundations of machine learning theory and data-driven sequential decision-making. His work has been recognized with a Best Paper Award at ICML ’24 and selection as a Rising Star in Data Science (University of California San Diego ’24). His postdoctoral research is supported by an NSF fellowship, and his Ph.D. studies were fully supported by the Israeli Council for Higher Education Scholarship for Outstanding PhD Students in Data Science.

Machine learning relies on its ability to generalize from limited data, yet a principled theoretical understanding of generalization remains incomplete. While binary classification is well understood in the classical PAC framework, even its natural extension to multiclass learning is substantially more challenging.

In this talk, I will present recent progress in multiclass learning that characterizes when generalization is possible and how much data is required, resolving a long-standing open problem on extending the Vapnik–Chervonenkis (VC) dimension beyond the binary setting. I will then turn to complementary results on efficient learning via boosting.  We extend boosting theory to multiclass classification, while maintaining computational and statistical efficiency even for unbounded label spaces.

Lastly, I will discuss generalization in sequential learning settings, where a learner interacts with an environment over time. We introduce a new framework that subsumes classically studied settings (bandits and statistical queries) together with a combinatorial parameter that bounds the number of interactions required for learning.

Bio: 

Nataly Brukhim is a postdoctoral researcher at the Institute for Advanced Study (IAS) and the Center for Discrete Mathematics and Theoretical Computer Science (DIMACS). She received her Ph.D. in Computer Science from Princeton University, where she was advised by Elad Hazan, and was a student researcher at Google AI Princeton. She earned her M.Sc. and B.Sc. in Computer Science from Tel Aviv University.

Recent years have seen a growing interest in developing coherent atom-light interfaces due to their relevance for cavity QED and quantum networks. The focus has been on systems with collectively coupled ensembles, and with strongly coupled single atoms. However, combining strong atom-light coupling with single-atom control remains challenging. We report on experiments with an atomic tweezer array that is strongly coupled to an optical fiber cavity. The setup integrates three ingredients: single atom control for arbitrary quantum logical operations, a tweezer-cavity system for cavity QED, and a direct fiber interface for real-time measurements and networking. This opens a path for programmable interactions within an atomic tweezer array, for non-destructive readout protocols, and for implementing quantum network protocols. 

Sample covariance matrices are among the most fundamental objects in random matrix theory and statistics. In this talk, I'll discuss recent work identifying the assumptions on random vectors that allow local laws to hold for their sample covariance matrices — these are matrices with iid rows sampled from a fixed distribution.

A local law says that the empirical eigenvalue distribution converges to its deterministic limit—in this case the deformed Marchenko–Pastur law—not just globally, but on short intervals which still contain a power of dimension many eigenvalues. This fine-grained control is essential for many applications, including universality for the local eigenvalue distributions.

The classical approach assumes the data vectors take a separable form g=Xw where w has independent entries—but this excludes many natural examples. We ask: what assumptions on g are really needed? It turns out that concentration of quadratic forms suffices for an optimal averaged local law, while a structural condition on cumulant tensors—interpolating between independence and generic dependence—suffices for the full anisotropic local law.

I'll discuss key examples where our assumptions can be verified: sign-invariant vectors, the 'random features model’ from machine learning, and some examples of spin-glass type. I'll also give a short overview of the proof, which introduces a tensor network framework for fluctuation averaging in the presence of higher-order cumulant structure. 

Joint with Jack Ma (Yale), Zhou Fan (Yale), Zhichao Wang (Berkeley)

Prof. Orr Spiegel from TAU studies animal movement

Two-dimensional materials are atomically thin crystals with a huge variety of physico-chemical properties. By stacking such materials into heterostructures we can combine the electrical, optical, and vibrational excitations of different materials with atomic control over their interfaces. Despite the great selection of 2D materials existing today, we desire novel routes for their preparation in addition to cleaving them from layered bulk parent compounds.In this talk I discuss a concept for novel 2D materials from organic molecules: Growing molecules into well-defined 2D and 1D lattices. We prepared 2D lattices of flat aromatic molecules using hexagonal boron nitride and graphene as atomically smooth substrates. The molecules are well separated in space and oriented side-by-side so that electrons and vibrations are confined to the individual building blocks. However, the interaction between their optical and vibrational transition dipole moments gives rise to collective states that can propagate inside the lattices. One-dimensional molecular lattices are grown by filling carbon- and boron-nitride nanotubes leading to giant J aggregates inside the tubes. We discuss how to use molecular lattices for advance molecular-2D-material heterostructures and how to manipulate their emergent optical excitations.