Online temporal pattern learning

摘要

This paper describes a biologically motivated approach, using hierarchical temporal memory (HTM), to build a high-level self-organizing visual system for a soccer bot. Meanwhile it presents two unsupervised online learning algorithms for temporal patterns in HTMs. The algorithms were implemented in a simulated soccer bot for a real-world evaluation. After a training phase, the robot was able to recognize different static objects in the soccer field. It also learned and recognized high-level objects that are composed of simpler objects, with position invariance and was also able to learn and recognize motions in the objects, all in a completely unsupervised manner.