Information Theory, Machine Learning, and Reproducing Kernel Hilbert Spaces

Abstract

The common problem faced by many data processing professionals is how to best extract the information contained in data. In our daily lives and in our professions, we are bombarded by huge amounts of data, but most often data are not our primary interest. Data hides, either in time structure or in spatial redundancy, important clues to answer the information-processing questions we pose. We are using the term information in the colloquial sense, and therefore it may mean different things to different people, which is OK for now. We all realize that the use of computers and the Web accelerated tremendously the accessibility and the amount of data being generated. Therefore the pressure to distill information from data will mount at an increasing pace in the future, and old ways of dealing with this problem will be forced to evolve and adapt to the new reality. To many (including the author) this represents nothing less than a paradigm shift, from hypothesis-based, to evidence-based science and it will affect the core design strategies in many disciplines including learning theory and adaptive systems.