Neural Architecture Search Enables Automated Rediscovery of Sparse Recovery Algorithms

Global: Neural Architecture Search Enables Automated Rediscovery of Sparse Recovery Algorithms

On December 25, 2025, researchers Patrick Yubeaton, Sarthak Gupta, M. Salman Asif, and Chinmay Hegde submitted a preprint to arXiv detailing a meta‑learning framework that leverages Neural Architecture Search (NAS) to automatically rediscover key components of the Iterative Shrinkage‑Thresholding Algorithm (ISTA) and its accelerated variant, Fast ISTA (FISTA). The work aims to streamline the traditionally heuristic‑driven process of designing algorithms for inverse problems in signal processing.

Background on Sparse Recovery

Inverse problems, such as signal deconvolution and compressed sensing, often rely on sparse recovery techniques like ISTA to obtain stable solutions. Designing and tuning these algorithms typically requires extensive domain expertise and iterative experimentation, which can impede rapid innovation.

Meta‑Learning Framework

The authors constructed a NAS‑based search space encompassing more than 50,000 variables, representing potential algorithmic components, hyperparameters, and update rules. By framing algorithm discovery as a differentiable optimization problem, the framework evaluates candidate architectures against a reconstruction loss, guiding the search toward effective sparse recovery strategies.

Results and Validation

Experimental results demonstrate that the framework successfully recovers essential elements of both ISTA and FISTA, including the proximal gradient step and momentum acceleration, without prior knowledge of the algorithms’ structures. Validation across multiple synthetic data distributions confirms the robustness of the discovered configurations.

Broader Applicability

Beyond ISTA/FISTA, the authors illustrate how the same NAS pipeline can be adapted to other inverse‑problem solvers and data regimes, suggesting a generalizable pathway for automated algorithm design in signal processing and related fields.

Future Directions

The study proposes extending the approach to larger‑scale problems, integrating hardware‑aware constraints, and exploring cross‑domain transfer of discovered algorithms. Such extensions could further reduce the manual effort required to develop high‑performance solvers for emerging applications.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.