Analog implementation of a permanent unsupervised learning algorithm

Abstract

Today, most applications of neural networks are devoted to pattern recognition in the fields of speech and image processing. The main connectionnist tools used for this purpose are associative memories [9], particularly multi-layer networks associated with the well-known algorithm of the backpropagation of the error gradient [10, 11, 13]. To obtain the best results with these devices, it is necessary to apply some preprocessing which increases the orthogonality degree of the prototypes to be stored. For instance, KOHONEN proposed such a preprocessing, very simple but efficient, with a Laplacian filter [9, pp. 170–71]. First, consider these prototypes, that is the input vectors of these memories. In fact, they come from the outputs of sensors (cameras, microphones or eyes, ears and so on…). These sensors are multisensitive: the signal provided by one sensor is a superimposition of signals emitted by all the sources of the neighbourhood. It is clear that decisions and classification of such a set of multidimensional data (prototypes) is difficult or even impossible, because of the redundancy of the components of prototype vectors associated to the mixtures. So, separating the independent sources in these mixtures is a very powerful preprocessing.