English

Algorithms and SQ Lower Bounds for Robustly Learning Real-valued Multi-index Models

Machine Learning 2025-05-28 v1 Data Structures and Algorithms

Abstract

We study the complexity of learning real-valued Multi-Index Models (MIMs) under the Gaussian distribution. A KK-MIM is a function f:RdRf:\mathbb{R}^d\to \mathbb{R} that depends only on the projection of its input onto a KK-dimensional subspace. We give a general algorithm for PAC learning a broad class of MIMs with respect to the square loss, even in the presence of adversarial label noise. Moreover, we establish a nearly matching Statistical Query (SQ) lower bound, providing evidence that the complexity of our algorithm is qualitatively optimal as a function of the dimension. Specifically, we consider the class of bounded variation MIMs with the property that degree at most mm distinguishing moments exist with respect to projections onto any subspace. In the presence of adversarial label noise, the complexity of our learning algorithm is dO(m)2poly(K/ϵ)d^{O(m)}2^{\mathrm{poly}(K/\epsilon)}. For the realizable and independent noise settings, our algorithm incurs complexity dO(m)2poly(K)(1/ϵ)O(K)d^{O(m)}2^{\mathrm{poly}(K)}(1/\epsilon)^{O(K)}. To complement our upper bound, we show that if for some subspace degree-mm distinguishing moments do not exist, then any SQ learner for the corresponding class of MIMs requires complexity dΩ(m)d^{\Omega(m)}. As an application, we give the first efficient learner for the class of positive-homogeneous LL-Lipschitz KK-MIMs. The resulting algorithm has complexity poly(d)2poly(KL/ϵ)\mathrm{poly}(d) 2^{\mathrm{poly}(KL/\epsilon)}. This gives a new PAC learning algorithm for Lipschitz homogeneous ReLU networks with complexity independent of the network size, removing the exponential dependence incurred in prior work.

Keywords

Cite

@article{arxiv.2505.21475,
  title  = {Algorithms and SQ Lower Bounds for Robustly Learning Real-valued Multi-index Models},
  author = {Ilias Diakonikolas and Giannis Iakovidis and Daniel M. Kane and Lisheng Ren},
  journal= {arXiv preprint arXiv:2505.21475},
  year   = {2025}
}
R2 v1 2026-07-01T02:43:50.412Z