Comparison of electrophysiological spatial pattern changes in short-and long-term learning

Abstract

In cognition, sensation - "the activity of a sense organ and closely connected nerve structures" - is followed by perception - "the meaningful impression of any object obtained by use of the senses" (Webster). Brains perform the microscopic sensory processes of extracting and processing information from the world. They have been well modeled with cellular neural networks (CNN) by modifying connections in reinforcement learning. Brains perform the mesoscopic perceptual processes of categorizing the information and constructing its meaning by using distributed memories. Perception requires a different topology than CNN, which is provided by random graph theory. Correspondingly cerebral cortex has two kinds of networks that learn. The sensory and motor cortices contain numerous, highly specialized, local networks that are tuned by local, rapid, short-term learning to select features in the input and bind them into sustained discharges of microscopic cell assemblies. These local systems are embedded in brain-wide, scale-free, mesoscopic feedback connections that learn slowly in consolidation. The distributed architecture integrates sensory input with selected memories into sequences of large-scale, distributed neural activity patterns that express meanings of sensory information. The patterns resemble cinematographic frames. The construction of each frame is by a phase transition. The brain topology, process, and activity patterns may be modeled by embedding multiple CNN in a large-scale random graph. That is neuropercolation; its goal is to provide a superior alternative to differential equations, which are now used to model perception.