Optimization Learning: Perspective, Method, and Applications

Risheng Liu

Introduction

Many fundamental tasks in computer vision and signal processing—such as image deblurring, denoising, and MRI reconstruction—are categorized as "inverse problems." These tasks involve recovering an original signal from corrupted observations. Historically, researchers have relied on two distinct paths: classical mathematical optimization, which is theoretically robust but often too rigid for complex data, and deep learning, which is highly flexible but often lacks formal convergence guarantees.

This paper introduces a unified framework that bridges this gap, providing a "best of both worlds" approach. By integrating learnable neural networks within a rigorous mathematical structure, the authors create an optimization paradigm that is both powerful and theoretically sound.

The General Optimization Learning Paradigm

The core contribution is a generic framework that allows researchers to embed neural networks into optimization iterations while maintaining strict convergence properties. As illustrated in the diagram below, the process involves translating domain knowledge into a structured optimization problem, which is then solved through a series of learnable stages.

The proposed framework translates domain knowledge into a structured optimization process, combining iterative stages with a knowledge-driven deep learning landscape.

Key Components

  • Abstract Iterative Modularization: The problem is decomposed into two learnable modules: $A_f$ (handling data fidelity) and $A_g$ (handling the image prior).
  • Optimality Error Checking: This acts as a "safety net." At each iteration, the algorithm monitors the error of the neural network's output to ensure it aligns with the mathematical objective.
  • Proximal Feedback Mechanism: If a network's output fails to meet specific criteria, the algorithm applies a proximal step to "correct" the trajectory. This ensures the objective function value consistently descends toward a solution.
  • Theoretical Convergence: The paper provides a formal proof (Theorem 1) demonstrating that this sequence globally converges to a critical point, even in non-convex and non-smooth scenarios.

Applications and Performance

The framework was evaluated across several challenging domains, including low-light enhancement, medical imaging (MRI), and image restoration.

Key Results

  • MRI Reconstruction: The method achieved a PSNR of 30.48, significantly outperforming traditional Total Variation (TV) methods (25.22) and existing deep unrolling models like ADMM-Net (27.91).
  • Image Deblurring: In non-blind deblurring tasks, the paradigm reached a PSNR/SSIM of 32.40 / 0.8935, surpassing established filters like EPLL.
  • Rain Streak Removal: The method effectively handles complex environmental noise, preserving fine textures while removing artifacts that often plague pure deep learning models.
Example of a real-world application: the framework is applied to tasks like rain streak removal to recover clear signals from noisy environmental data.

Insights and Future Directions

The primary insight of this work is that domain knowledge (mathematical models) and data-driven learning (AI) are not mutually exclusive. By using math to define the structure of a solution and AI to learn the parameters of that structure, we can create models that are safe for mission-critical applications like medical imaging.

While the current framework relies on the objective function being "semi-algebraic" for the strongest convergence results, future work aims to extend this paradigm to "blind" inverse problems, where the specific corruption process (such as a blur kernel) is unknown at the outset. This flexibility makes the framework a powerful "wrapper" capable of upgrading existing heuristic deep models into theoretically guaranteed optimization tools.

Summary updated

BibTeX
@inproceedings{liu2020optimization,
  author = {Risheng Liu},
  title = {{Optimization Learning: Perspective, Method, and Applications}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.24963/ijcai.2020/728},
}
Figure