Greedy Learning to Optimize with Convergence Guarantees

TL;DR

This paper introduces a greedy learning method for optimization that allows training over many iterations with guaranteed convergence and improved empirical performance on inverse problems.

Contribution

A novel greedy strategy for learning iteration-specific parameters that ensures convergence guarantees and enables training over many more iterations with constant memory.

Findings

Convergence guarantees for learned algorithms on training and unseen functions.

Improved empirical performance on image deblurring and CT reconstruction.

Effective convolutional preconditioners outperform classical methods.

Abstract

Learning to optimize is an approach that leverages training data to accelerate the solution of optimization problems. Many approaches use unrolling to parametrize the update step and learn optimal parameters. Although L2O has shown empirical advantages over classical optimization algorithms, memory restrictions often greatly limit the unroll length and learned algorithms usually do not provide convergence guarantees. In contrast, we introduce a novel method employing a greedy strategy that learns iteration-specific parameters by minimizing the function value at the next iteration. This enables training over significantly more iterations while maintaining constant device memory usage. We parameterize the update such that parameter learning is convex when the objective function is convex. In particular, we explore preconditioned gradient descent and an extension of Polyak's Heavy Ball Method with multiple parametrizations including a novel convolutional preconditioner. With our learned algorithms, convergence in the training set is proved even when the preconditioners are not necessarily symmetric nor positive definite. Convergence on a class of unseen functions is also obtained under certain assumptions, ensuring robust performance and generalization beyond the training data. We test our learned algorithms on two inverse problems, image deblurring and Computed Tomography, on which learned convolutional preconditioners demonstrate improved empirical performance over classical optimization algorithms such as Nesterov's Accelerated Gradient Method and the quasi-Newton method L-BFGS.