Guest Editorial: Machine Learning for Signal Processing

Abstract

One of the fundamental goals of signal processing is to model observed signals or their underlying processes accurately. This is a very challenging problem due to issues such as observation noise and the finite nature of observations. To overcome these difficulties, signal processing has proactively employed the relatively new and powerful discipline of machine learning. The type of signal processing that incorporates machine learning as its booster, i.e., Machine Learning for Signal Processing (MSLP), is now working as a promising technological paradigm that opens the way to discovering salient features from noisy signals, predicting unseen signals, and achieving accurate modeling of the process that generates signals. A large number of papers related to MLSP have been published in different places. However, to advance research in the area effectively, one comprehensive forum would clearly be advantageous. Motivated by this need, we organized the publication project of this special issue onMSLP. Twenty-one papers were submitted to the issue, and the eleven papers presented here were selected through a rigorous review process. The selected papers cover most of the key factors of signal processing, in particular its important sub-field of pattern classification. Pattern classification basically consists of a featureextraction (selection) stage, which converts an input pattern to features, and a classification stage, which labels a set of converted features as one of various preset class indexes. Research on pattern classification sometimes focuses on either of these component stages. Each of the selected papers more or less refers to both of the stages, but they can be divided into three groups to give readers a clear understanding of the issue and the relative positions of the included papers. The first four papers focus on theoretical or algorithmic research topics related to the classification stage by studying ways of designing (training) class models. The next three papers study the extraction of salient features, or those robust to unseen patterns, from a limited amount of input data. Finally, the remaining four papers study applications of signal modeling and/or pattern classification to particular types of input data, such as face images and speech signals.