Ada-LISTA: Learned Solvers Adaptive to Varying Models.

Area of Science:

• Machine Learning
• Signal Processing
• Computer Vision

Background:

• Learned Iterative Soft Shrinkage (LISTA) networks accelerate performance but require fixed dictionaries.
• This limitation restricts their applicability in dynamic or uncertain model scenarios.
• Non-learned iterative solvers offer flexibility but lack the speed of learned approaches.

Purpose of the Study:

• To introduce Ada-LISTA, an adaptive learned solver with a universal architecture.
• To enable efficient sparse coding under varying signal and dictionary models.
• To demonstrate the adaptability and application of the proposed scheme.

Main Methods:

• Developed Ada-LISTA, an adaptive learned iterative solver.
• Input signal and corresponding dictionary are used to train a universal architecture.
• Theoretical and numerical analysis was conducted to validate the approach.

Main Results:

• Ada-LISTA achieves linear convergence rates for sparse coding.
• The architecture adapts to permutations and perturbations of the dictionary.
• Demonstrated successful application in natural image inpainting tasks.

Conclusions:

• Ada-LISTA provides a flexible and efficient solution for sparse coding.
• The adaptive nature overcomes limitations of fixed-dictionary learned solvers.
• This approach enhances applicability in diverse real-world scenarios, including image restoration.