Cooperation of neural nets and task decomposition

In order to solve classification problems when using neural nets, it may be necessary to decompose the global task into a few subtasks, each being processed by one specialized network or module. The task is thus performed through a cooperation between all the modules. Such a decomposition may help to find a more efficient solution. The authors describe experiments on cooperating nets, in this case a task that can be totally decomposed by the user using a priori knowledge of the problem. The example considered clearly shows that building an architecture made of independent specialized units is not the most efficient way to perform successive processings of the data. Important improvements, both in classification and speed performances, may be obtained by training the different modules together. Another important point is that this modularity is the only way to build architectures which are suitable for complex tasks