ICALA: Incremental clustering and associative learning architecture

I propose a learning and memory architecture which can incrementally learn and associate an increasing number of patterns. The approach consists of the integration of two methods - a topology learning algorithm to perform incremental clustering, and an associative memory model to learn relationships based on the co-occurrence of input patterns. The approach supports online learning, is tolerant to noise, and generally applicable to any kind of real-valued vector data. I tested the proposed architecture on an incremental associative learning task with visual patterns. Evaluations were performed both in a simulated setup and with a real robot. Results showed that the architecture could learn nearly all presented patterns but in some cases the recall rate decreased as these patterns were retrieved. I suggest ways to overcome this effect and also discuss future work aimed at achieving a better performance.