Bi-directional Modularity to Learn Visual Servoing Tasks

This paper shows the advantage of using neural network modularity over conventional learning schemes to approximate complex functions. Indeed, it is difficult for artificial neural networks like Kohonen extended maps to converge toward an efficient and adequate solution when the dimensionality of the input and output spaces are high. Associated to an appropriate learning technique, modularity in neural networks is able to overcome the high dimensionality of the input/input space by decomposing it into different intermediate spaces of reduced dimensionality. The decomposition results in independent neural modules. The efficiency of this learning technique will be enlightened with a visual servoing application. In this application, the relationship between the visual features issued from a stereoscopic vision system and the angles of a 5 DOF-robot will be learned and approximated. Simulations have been conducted and clearly show that this complex, nonlinear, and high dimensional function can be learned efficiently with the neural network modularity approach. Moreover, we show through these simulations that neural modules can be re-utilized, thus reducing the convergence time of the learning and the memory requirements.