Neural Object Recognition by Hierarchical Learning and Extraction of Essential Shapes

We present a hierarchical system for object recognition that models neural mechanisms of visual processing identified in the mammalian ventral stream. The system is composed of neural units organized in a hierarchy of layers with 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 pure visual experience and with no prior labeling. Object labels are only introduced to train a classifier on the system’s output. The system’s tuning takes place in a feedforward path. We also present a neural mechanism for back projection of the learned image patterns down the hierarchical layers. This feedback mechanism could serve as a starting point for integration of whatand where-information processed by the ventral and dorsal stream. We test the neural system with natural images from publicly available datasets of natural scenes and handwritten digits.