Neural network super architectures

The influence of the initial architecture of a neural network on its learning ability is examined. Various architectural traits are combined in order to form a set of permutations of architectures each of which is trained for a benchmark problem. The results are compared to the original architectures of the benchmark problem. It can be observed from the results of the experiments that there exists a set of suitable architectures for each problem and in these sets certain ´super´ architectures can be identified which produce the best result for the problem.