480. Adaptively Inexact Bilevel Learning via Primal-Dual Differentiation
Invited abstract in session MC-3: First-order methods in modern optimization (Part II), stream Large scale optimization: methods and algorithms.
Monday, 14:00-16:00Room: B100/4011
Authors (first author is the speaker)
| 1. | Mohammad Sadegh Salehi
|
| Independent Researcher |
Abstract
Bilevel optimisation plays a crucial role in machine learning, particularly for tasks such as learning parameters in variational regularisation problems. In this talk, we introduce an Adaptively Inexact First-order Method for Bilevel Learning, which employs a primal-dual differentiation strategy and leverages the ‘piggyback’ algorithm to compute hypergradients. Our approach derives an a posteriori error bound for the primal-dual setting, enabling adaptive tolerance selection to balance computational efficiency and accuracy. Additionally, we introduce an adaptive step-size strategy to enhance upper-level optimisation.
From an application perspective, we showcase the efficient learning of total variation discretisation for image denoising and the first successful bilevel learning of input convex neural networks (ICNNs) as regularisers via reformulation for sparse-angle computed tomography. Moreover, our method outperforms existing approaches for training such data-adaptive regularisers.
Keywords
- Multi-level optimization
- Large-scale optimization
- First-order optimization
Status: accepted
Back to the list of papers