NEURAL OBJECT RECOGNITION BY HIERARCHICAL APPEARANCE LEARNING

Abstract

We present a system for object recognition that is largely inspired by physiologically identified processing streams in the visual cortex, specifically in the ventral stream. It consists of neural units organized in a hierarchy of layers with encoding features of increasing complexity. A key feature of the system is that the neural units learn their preferred patterns from visual input alone. Through this "soft wiring" of neural units the system becomes tuned for target object classes through passive visual experience and no labels are required in this stage. Object labels are only introduced in the last step to train a classifier on the system's output. While this tuning process is purely feed-forward we also present a neural mechanism for back projection of the learned image patterns down the hierarchical layers, and demonstrate how this feedback can be used to stabilize the system in the presence of noise. We test the neural system with natural images from publicly available data-sets of natural scenes and handwritten digits.