EUROPT 2025
Abstract Submission

458. Optimization Techniques for Learning Multi-Index Models

Invited abstract in session MC-3: First-order methods in modern optimization (Part II), stream Large scale optimization: methods and algorithms.

Monday, 14:00-16:00
Room: B100/4011

Authors (first author is the speaker)

1. Hippolyte Labarrière
Università di Genova
2. Shuo Huang
DIBRIS - Università di Genova
3. Ernesto De Vito
DIMA - Università di Genova
4. Lorenzo Rosasco
MaLGa, DIBRIS, Università di Genova
5. Tomaso Poggio
McDermott Professor in Brain Sciences, MIT

Abstract

Neural networks naturally identify low-dimensional structures that improve generalization, a property often lacking in classical kernel methods. In this talk, I will present how we can bridge this gap by integrating the multi-index model (MIM) into kernel regression. By introducing a hyper-kernel that incorporates a learned projection matrix, we enhance the adaptability of kernel methods. I will discuss the theoretical guarantees of this approach and introduce an optimization framework based on alternating minimization techniques and Nyström approximations

Keywords

Status: accepted


Back to the list of papers