ART FCMAC: a memory efficient neural network for robotic pose estimation

The feature cerebellar model arithmetic computer (FCMAC) is a multiple-input-single-output neural network which can provide three-degree-of-freedom (3-DOF) pose estimation for a robotic vision system. In this paper a new architecture is introduced which combines the FCMAC with an adaptive resonance theory (ART) network. The ART module clusters patterns observed during training into a set of prototypes that are used to build the FCMAC. As a result, the network no longer grows without bound, but constrains itself to a user-specified size. Pose estimates remain accurate since the ART tends to discard the least relevant information first. In some cases the smaller network is better for generalization, resulting in a reduction of error at recall time. The ART-C algorithm is extended to include initial filling with randomly selected patterns (referred to as ART-F). In experiments using a real-world data set, the new network performed equally well using less than one tenth the number of coarse patterns as a regular FCMAC. In validation experiments, the FCMAC system outperformed radial basis function (RBF) networks for the 3-DOF problem, and had comparable performance to principle component analysis (PCA) which estimates orientation only.