Modularity in Artificial Neural Networks

Background

When designing a controller for an agent, there exists two main approaches: a single module contains all the agent’s required behaviours (monolithic approach), or global behaviour is decomposed into a set of simpler sub-behaviours, each one implemented by one module (modular approach). Monolithic controllers implement on a single module all the required mappings between the agent’s inputs and outputs. As an advantage, it is not required to identify required sub-behaviours nor relations between them. As a drawback, whether the complexity of the controller is high, it could be impossible at practice to design such a controller without obtaining large interferences between different parts of it. Instead, when a modular controller is used, the global controller is designed by a group of sub-controllers, so required sub-controllers and their interactions for generating the final global output must be defined.

Despite the disadvantages of the modular approach (Boers, 1992), complex behaviour cannot be achieved without some degree of modularity (Azam, 2000). Modular controllers allow the acquisition of new knowledge without forgetting previously acquired one, which represents a big problem for monolithic controllers when the number of required knowledge rules to be learned is large (De Jong et al., 2004). They also minimize the effects of the credit assignment problem, where the learning mechanism must provide a learning signal based on the current performance of the controller. This learning signal must be used to modify the controller parameters which will improve the controller behaviour. In large controllers, it becomes difficult finding changing parameters of the controller based on the global learning signal. Modularization helps to keep small the controllers’ size, minimizing the effect of the credit assignment.

Modular approaches allow for a complexity reduction of the task to be solved (De Jong et al., 2004). While in a monolithic system the optimization of variables is performed at the same time, resulting in a large optimization space, in modular systems, optimization is performed independently for each module resulting on reduced searching spaces. Modular systems are scalable, in the sense that former modules can be used for the generation of new ones when problems are more complex, or just new modules can be added to the already existing ones. It also implies that modular systems are robust, since the damage on one module results in a loss of the abilities given by that module, but the whole system is partially kept functioning. Modularity can be a solution to the problem of neural interference (Di Ferdinando et al., 2000), which is encountered in monolithic networks. This phenomenon is produced when an already trained network losses part of its knowledge when either, it is re-trained to perform a different task, called temporal cross-talk (Jacobs et al.,1991), or two or more different tasks at the same time, the effect being called spatial cross-talk (Jacobs,1990). Modular systems allow reusing modules in different activities, without re-implementation of the function represented on each different task (De Jong et al., 2004) (Garibay et al., 2004).