An adaptive resonance theory-based neural network capable of learning via representational redescription

This paper introduces a neural network architecture called R2MAP, which is based on the representational redescription hypothesis in cognitive science and adaptive resonance theory (ART) neural networks. The R2MAP network learns to classify arbitrary sequences of input patterns using a re-iterative process whereby knowledge that gets embedded in the network via ARTMAP-style error-driven learning is redescribed and becomes available to it for further learning. The knowledge redescription phase is triggered when the perceived level of difficulty of the given task exceeds a certain threshold, and is achieved through the dynamic creation of new features that better distinguish between output classes. This way the R2MAP network is capable of learning complex, relational input-output dependencies that cannot be represented efficiently using solely the features extracted through ordinary learning of statistical relationships. A simple proof-of-concept example is presented to illustrate the main ideas. Some related work is also discussed