About some perception problems in neural networks

Abstract

Perception is a major problem which is studied as well in life science as in engineering science: This paper concerns a reflection about some of the early mechanisms which underlie perception. Examples are taken in the field of vision in biology and in computer vision, showing the necessity of some adequate pre-processing of the signals. Then, perception appears as a process of representation of signals, that is, as a process of data analysis aimed at finding the structure of the data. Two examples of artificial neural networks are presented to illustrate the problem of data representation. The first one called “Independent Component Analysis” is close to the signal level, the second one, called “Curvilinear Component Analysis” can be seen as a smooth transition between the aspects of Signal Processing and those of Data Analysis.