A visual novelty detection filter based on bag-of-words and biologically-inspired networks

The ability of a robot system to learn continuously until the end of its cycle is a desired feature and consists a difficult challenge for the robotics research community. One of the main components necessary for effective continuous learning is the behaviour of a robot of identifying and focusing its attention to novel patterns, and has been an active area of research over the last decade, considering the large number of surveys that have been published recently. This paper presents the initial steps of a larger work which is concerned with continuous learning driven by novelty detection as an intrinsic motivation. For the learning structure and the novelty detection filter we use a bag-of-words model combined with unsupervised biologically inspired neural networks, both for the generation of the vocabulary and the learner/classifier.