A TAXONOMIC ASSOCIATIVE MEMORY-BASED ON NEURAL COMPUTATION

A single layer feed forward neural network for associating organisms w ith binary features to the most typical organisms in a given numerical classification is presented. The network implements an associative me mory, which has stored maximal predictivity. The network also represen ts a neural model for the classification as well as a neurocomputer fo r numerical identification. The rationale in probabilistic numerical i dentification of bacteria is explained. After a learning phase based o n backpropagation for minimization of the corssentropy between the mos t typical organisms and the network outputs the memory associates by m aximizing a kind of 'probability of belonging' to a taxon. For numeric al experiments we have modified the MATLAB(TM)Neural Networks Toolbox. We consider in particular the expansion and rejection of identificati on properties of the memory, which are potentially useful in cumulativ e or continuous classification and identification.