Interpretable, efficient, and dependable signal processing algorithms via deep unfolding

CPU and GPU frequencies hit physical limitations around 2005 and have remained stagnant. Meanwhile, transistor counts continued to double every two years following Moore’s law.

This has led to a shift towards parallel computation as the default paradigm for meeting today’s computational demands. However, many signal processing algorithms based on numerical optimization are still essentially serial.

We aim to change this by providing tools, best practices, and an analytical framework to parallelize these algorithms.We will achieve this via two key technologies: proximal optimization and deep unfolding.

Proximal optimization is a class of numerical optimization algorithms suitable for solving large-scale optimization problems as they can naturally decompose problems for parallel computations.

Deep unfolding is the idea of interpreting the iterations of an optimization algorithm as layers of an artificial neural network, making them amendable to be run on hardware designed for deep learning.

By combining these technologies, we aim to improve the efficiency of the algorithms when implemented on massively parallel hardware and provide guarantees on reliability and generalizability by ensuring that fundamental properties of the underlying optimization algorithms, such as convergence, are retained by the neural network structure.