DiCoDiLe: Distributed Convolutional Dictionary Learning.

Area of Science:

• Signal Processing
• Machine Learning
• Computer Vision

Background:

• Convolutional Dictionary Learning (CDL) learns shift-invariant patterns for signal and image representation.
• Existing CDL optimization methods struggle with extremely large datasets due to computational and memory constraints.
• CDL is valuable for applications like image denoising, inpainting, and multivariate signal pattern discovery.

Purpose of the Study:

• To propose a novel distributed and asynchronous algorithm for Convolutional Dictionary Learning (CDL).
• To address the computational challenges of optimizing CDL on massive datasets.
• To enable CDL pattern recovery and learning on data exceeding single-computer memory.

Main Methods:

• Developed a distributed and asynchronous algorithm utilizing locally greedy coordinate descent.
• Implemented a soft-locking mechanism, eliminating the need for a central server.
• Algorithm computation scales linearly with data size, distributing workload across workers.

Main Results:

• Experimental results confirm the algorithm's theoretical linear scaling properties.
• Demonstrated improved pattern recovery for increasingly large images.
• Successfully learned patterns from Hubble Space Telescope images with tens of millions of pixels.

Conclusions:

• The proposed distributed CDL algorithm effectively overcomes computational and memory limitations of large-scale data.
• The method enables efficient and scalable pattern discovery and recovery in massive image and signal datasets.
• This approach opens possibilities for analyzing unprecedentedly large scientific imaging data.