A novel network classifier and its application: optimal hierarchical orderings of the cat visual system from anatomical data

In the last few years a large amount of data has been collected on the anatomical structure of the cat visual system. The system is composed of many distinct areas, each of which is interconnected with many others. These anatomical data reflect the fact that the system is formidably complex, and their complexity has prompted the search for effective methods of data analysis. Connections between visual cortical areas can be classified as 'ascending', 'descending' or 'lateral' according to the pattern of their terminations and origins within the cortical layers. This categorisation has led to 63 pairwise hierarchical relations for 22 visual areas. Some of the hierarchical constraints are, however, inconsistent with others and the presently available anatomical data are by no means complete. For these reasons, and since the number of possible orderings for 22 areas is very large, we analysed the overall structure of the data with the help of an optimisation procedure. This procedure employed a modified annealing algorithm, which was integrated into a network processing environment, CANTOR, and searched for optimal hierarchical orderings, that is, those with a minimal number of hierarchical constraint violations. Starting from an arbitrarily chosen structure, it proceeded by cumulative modification and subsequent cost evaluation of candidate solutions.