Learning the Unknown A Computational Approach to Incrementally Developing a Hypothesis Space with Gaussian Mixture Models

Abstract

We present a new computational model that captures a key behaviour found in humans we wish to replicate, namely the ignoring of singular anomalous stimuli. Our model is built around incremental learning in Gaussian mixture models combined with an explicit `unknown' hypothesis with posterior reasoning. We show through simula- tions that our model indeed captures the behaviour of interest as opposed to IGMM, the model it was based on. We also show that our model improves the performance of IGMM, providing theoretical evidence that ignoring anomalies is a superior method of learning as opposed to tting all the data.