An Adaptive Resonance Theory-based Neural Network for Autonomous Learning via Iterative Knowledge Redescription

This paper introduces a neural network architecture called ARTIR, which autonomously learns classifications of arbitrary input sequences via iterative redescription of knowledge that it builds up during learning. Knowledge redescription is achieved via changes to both the network architecture and the connection weights between neurons. Based on ARTMAP, the supervised version of the Adaptive Resonance Theory (ART) neural network, and motivated by the Representational Redescription hypothesis in cognitive science, the ARTIR network creates new, successively deeper internal features iteratively. The process is driven by the complexity of the classification task as measured by the number of input categories the ARTMAP network develops during training. The network is forced to discover a new internal feature that better discriminates the output classes every time the currently perceived difficulty of the learning task exceeds a certain limit as indicated by the number of input categories the ARTMAP network develops in response to an input-target sequence. Once a new candidate feature - found to correlate well enough with an output class - is accepted, ARTIR re-learns the mapping using the additional feature and uncommits its categories that are no longer needed as a result of re-learning. The iterative learning process then continues where the newly created feature, along with all original inputs and internal features created during previous stages of learning, becomes subject of internal knowledge redescription for future iterations. Experiments carried out on two binary decision problems demonstrate the viability of the approach, and the results point to further questions that will be the subject of future research.