Events

In multiclass classification with bandit feedback, a learner only observes whether its predicted label was correct, rather than the true label itself.  This simple change turns a basic supervised learning problem into a much harder exploration problem: how can one learn accurately over a large set of possible labels while receiving only minimal feedback?  In this talk, I will discuss a recent line of work studying the statistical complexity of bandit multiclass classification.  Perhaps surprisingly, we show that bandit feedback can be essentially free: with the right exploration strategy, one can match the full-information sample complexity rates up to logarithmic factors. The key technical ingredient is a new algorithmic approach to low-variance exploration that scales with the sparsity of the rewards, rather than with their ambient dimension.  Beyond resolving a substantial gap in prior work, these ideas also extend naturally to richer prediction problems such as list classification, combinatorial feedback models, and contextual bandits; somewhat unexpectedly, they have also played a role in resolving the tight sample complexity of standard full-information multiclass classification.

Based on joint work with PhD student Liad Erez and collaborators Fan Chen, Alon Cohen, Steve Hanneke, Yishay Mansour, Shay Moran, Sasha Rakhlin, and Qian Zhang.

In multiclass classification with bandit feedback, a learner only observes whether its predicted label was correct, rather than the true label itself.  This simple change turns a basic supervised learning problem into a much harder exploration problem: how can one learn accurately over a large set of possible labels while receiving only minimal feedback?  In this talk, I will discuss a recent line of work studying the statistical complexity of bandit multiclass classification.  Perhaps surprisingly, we show that bandit feedback can be essentially free: with the right exploration strategy, one can match the full-information sample complexity rates up to logarithmic factors. The key technical ingredient is a new algorithmic approach to low-variance exploration that scales with the sparsity of the rewards, rather than with their ambient dimension.  Beyond resolving a substantial gap in prior work, these ideas also extend naturally to richer prediction problems such as list classification, combinatorial feedback models, and contextual bandits; somewhat unexpectedly, they have also played a role in resolving the tight sample complexity of standard full-information multiclass classification.

Based on joint work with PhD student Liad Erez and collaborators Fan Chen, Alon Cohen, Steve Hanneke, Yishay Mansour, Shay Moran, Sasha Rakhlin, and Qian Zhang.

Queen Victoria charmed by Carroll’s novel, let it be known to the author that she would be pleased to receive a copy of his forthcoming work. Lewis Carroll was rumoured to have sent her his next volume having the enticing title: “An elementary treatise on determinants”. Thereby Queen Victoria could have acquired practically all necessary to understand the work presented here.

In this lecture we describe the components of an exponentially increasing family of varieties coming from invariant theory.

In more detail, let B be the subgroup of upper triangular matrices in GL(n, ℂ). Then a parabolic subgroup P containing B is given by a composition (c₁, c₂, …, cₖ) of n whose entries are the sizes of Levi factors of P. These are formed from the c_i × c_i blocks down the diagonal, whilst the Lie algebra 𝔪 of the nilradical of P is the subspace lying strictly above these blocks.

Then the variety whose components we seek is the nilfibre N for the adjoint action of P on 𝔪, for all compositions of n and all n. It is the zero locus of the Benlolo–Sanderson generators of the semi-invariant ring ℂ[𝔪]^P′, which are g in number.

We define [4] “The Red Set” which can be viewed as “connected” subsets of the diagonal entries in each Levi block for which just the lowest entry may have multiplicity > 1.

We construct [2] a “component map” from the Red Set to the set of irreducible components of P and show it to be injective. For this one only needs to know how to multiply out a determinant and some very modest intersection theory in complex projective space.

Successive linearisation of invariants gives the components obtained from the component map, but may eliminate some potential ones, since it involves inhomogeneous substitution.

Recently Fittouhi [1] has switched the emphasis to the factorisation coming from equal height columns in a reverse tableau [3]. For a component of dimension ≤ dim V − g, descent on Krull dimension shows that this component lies in one coming from the image of the component map. Alternatively one may recognize that this factorisation only requires excluded root vectors. These involve only homogeneous substitution, so no potential components are eliminated.

While the latest diffusion models excel at image editing, achieving precise, fine-grained control remains challenging due to entangled semantic concepts. This talk presents two recent papers introducing mechanisms for continuous and disentangled control over diffusion model conditioning. By mapping and isolating specific semantic dimensions, we can decouple entangled concepts. This enables highly targeted image editing that preserves the original identity while unlocking a new level of precision in generative models.]

Bio:

Daniel Garibi is a Ph.D. student in Computer Science at Tel-Aviv University, supervised by Daniel Cohen-Or and Hadar Averbuch-Elor. His research interests include personalization and precise image editing. Daniel is currently interning at Meta.