Emerging Machine Learning Techniques in Signal Processing

In the era of knowledge-based society and machine automation , there is a strong interest in machine learning (ML) techniques in a wide range of applications. The attention paid to ML methods within the DSP community is not new. Speech recognition is an example of an area where DSP and machine learning have been combined to develop efficient and robust speech recognizers. Channel equalization is another area at the intersection of ML and DSP techniques. After all, deciding upon the transmitted information symbol is nothing but a class assignment task. In cognitive radio, DSP techniques and ML methods can work together for developing algorithms for the efficient utilization of the radio spectrum. Image/video/audio coding, recognition, and retrieval are some additional typical examples where DSP and ML tie together. Another problem at the heart of the DSP community interests is the regression task, which can be cast as an ML problem. Biomedical applications constitute another area in which mixed ML and DSP ideas proved to be useful. Over the past years, a number of new powerful ML techniques have been developed, which are suitable for nonlinear processing and the general case of non-Gaussian data, and also for nondifferentiable cost functions or cost functions referring to robust statistics. Adaptive versions of some of these techniques have only recently started being studied. This is an area in which the DSP community has a lot to say and contribute. The focus of this special issue is twofold: (a) to consider novel theoretical results in ML methods and algorithms in the light of typical DSP applications, and (b) to report novel results obtained by the application of ML techniques in some typical DSP tasks. The special issue consists of 14 accepted papers. The papers can be divided into two main categories: (a) theory and algorithms and (b) applications. The latter covers a rather broad range of areas, such as communications, audio processing and recognition, medicine and neuroscience, and implementation architectures. The paper authored by H. Li and T. Adali extends previous works on using complex-valued calculus in order to implement nonlinear adaptive signal processing algorithms in the complex domain, as well as some procedures resulting from this framework. The paper authored by W. Liu and J. C. Príncipe presents a general framework for using the " kernel trick " in the context of affine projection algorithms, as well as some procedures resulting from this framework. …