Hierarchical Models of the Visual System

Abstract

Hierarchical models of the visual system are neural networks with a layered topology: In these networks, the receptive fields of units (i.e., the region of the visual space that units respond to) at one level of the hierarchy are constructed by combining inputs from units at a lower level. After a few processing stages, small receptive fields tuned to simple stimuli get combined to form larger receptive fields tuned to more complex stimuli. Such anatomical and functional hierarchical architecture is a hallmark of the organization of the visual system. Since the pioneering work of Hubel and Wiesel (1962), a variety of hierarchical models have been described from relatively small-scale models of the primary visual cortex to very large-scale (system-level) models of object and action recognition, which account for processing in large portions of the visual field and entire visual streams. The term ‘model of the visual system’ is generally reserved for architectures that are constrained in some way by the anatomy and the physiology of the visual system (with various degrees of realism). Convolutional networks are closely related connectionist networks with a similar architecture that have been used in multiple real-world machine learning problems including speech and music classification.