Use Online Dictionary Learning to Get Parts-Based Decomposition of Noisy Data

Abstract

A huge amount of data is generated every day. Extracting interpretable features from the data is becoming important. Meanwhile, dimension reduction and low-rank approximation are also becoming important as people want to factorize big matrix into smaller ones that are easy to handle. Sparse coding is such a technique that can factorize a matrix into sparse linear combinations of basis elements. We found that through Online Dictionary Learning, an efficient sparse coding method, we can decompose large data matrix with noise into interpretable dictionary atoms. Such atoms are useful in reconstructing a denoised data matrix.