Distinctive features should be learned

Abstract

Most existing machine vision systems perform recognition based on a fixed set of hand-crafted features, geometric models, or eigen-subspace decomposition. Drawing from psychology, neuroscience and intuition, we show that certain aspects of human performance in visual discrimination cannot be explained by any of these techniques. We argue that many practical recognition tasks for artificial vision systems operating under uncontrolled conditions critically depend on incremental learning. Loosely motivated by visuocortical processing, we present feature representations and learning methods that perform biologically plausible functions. The paper concludes with experimental results generated by our method.