Core Concepts
The authors propose a novel subspace-based framework, LancBiO, that leverages the Krylov subspace and the Lanczos process to efficiently and accurately approximate the Hessian inverse vector product in bilevel optimization problems.
Abstract
The paper addresses the computational challenges in bilevel optimization, where the calculation of the hyper-gradient involves a Hessian inverse vector product, which is a bottleneck. To circumvent this, the authors construct a sequence of low-dimensional approximate Krylov subspaces using the Lanczos process. This allows for dynamically and incrementally approximating the Hessian inverse vector product with less effort, leading to a favorable estimate of the hyper-gradient.
The key aspects of the proposed LancBiO framework are:
Dynamic Krylov subspace construction: LancBiO builds up a dynamic process for constructing low-dimensional subspaces tailored from the Krylov subspace. This reduces the large-scale subproblem to a small-size tridiagonal linear system, drawing on the Lanczos process.
Incremental Hessian inverse approximation: The constructed subspaces enable dynamically and incrementally approximating the Hessian inverse vector product across outer iterations, thereby enhancing the estimate of the hyper-gradient.
Restart mechanism and residual minimization: LancBiO incorporates a restart mechanism to mitigate the accumulation of differences in the Hessian matrices, and solves a residual minimization subproblem to leverage historical information and improve the approximation accuracy.
The authors provide theoretical analysis to show the global convergence of LancBiO with an O(ε^-1) rate. Experiments on a synthetic problem and two deep learning tasks demonstrate the efficiency and effectiveness of the proposed approach compared to existing bilevel optimization methods.
Stats
The authors do not provide any specific numerical data or statistics in the content. The content focuses on the algorithmic development and theoretical analysis of the proposed LancBiO framework.