Validity Index-Based Vigilance Test in Adaptive Resonance Theory Neural Networks

One of the distinguishing features of Adaptive Resonance Theory (ART) is that it relies on a second similarity check, called a vigilance test, to accept or reject a sample into a given category. Generic unsupervised versions of ART rely on a single layer vigilance test, whereas their supervised counterparts possess a second layer test based on classification errors that trigger a match tracking procedure regulated by an inter-ART block. This work uses a second layer vigilance test based on validity indices. A new sample is accepted into a category if its match function surpasses the vigilance test of both layers: the standard first check is based on minimum similarity, and the second check analyses whether setting that sample as belonging to the winner category results in an improvement of the current data partition according to the chosen validity index used as a cost function. Namely, if the new clustering state is superior to the previous one, then learning is allowed for the winning category. Otherwise, the algorithm proceeds as usual in ART implementations. Thus, this local greedy heuristic uses the validity index as a reinforcement signal, looking at the immediate reward to guide the learning of the ART categories without an additional external optimizer algorithm. A sweep analysis of the first layer vigilance parameter was performed and experiments indicate that the presented approach outperforms the standard Fuzzy ART neural network when samples are randomly presented. When samples are presented in a predefined order, Fuzzy ART obtains the best peak performance, however the modified approach was less sensitive to parameter variations.