Brain-Structured Connectionist Networks That Perceive and Learn

This paper specifies the main features of Brain-like, Neuronal, and Connectionist models; argues for the need for, and usefulness of, structuring networks of neuron-like units into successively larger brain-like modules; and examines Recognition Cone models of perception from this perspective, as examples of such structures. Neuroanatomical, neurophysiological, and behavioral data on the structure, function, and development of the visual system are briefly summarized to motivate the architecture of brain-structured networks for perceptual recognition. The structural and functional architecture of Recognition Cones, the flow of information and the parallel-distributed nature of processing and control in Recognition Cones are described. The results from the simulation of carefully designed Recognition Cone structures that perceive objects (e.g., houses) in digitized photographs are presented. A framework for perceptual learning, including mechanisms for generation-discovery, that involves feedback-guided growth of new links between neuron-like units as needed, within a dynamically emerging network topology, subject to brain-like constraints on the network connectivity (e.g., local receptive fields, global convergence-divergence) is introduced. The information processing transforms discovered through generation are fine-tuned by feedback-guided reweighting of links. A case is made for the need for generation and discarding of transforms in addition to reweighting of links in Connectionist networks for perceptual learning. Some preliminary results from the simulation of brain-structured networks that learn to recognize simple objects (e.g., letters of the alphabet, cups, apples, bananas) through feedback-guided generation and reweighting of transforms are presented. Experimental comparisons indicate that such networks can give large improvements over networks that either lack brain-like structure or/and learn by reweighting of links alone. The role of brain-like structures and generation in perceptual learning is examined. Some directions for future research are outlined