An analysis of Hebb's cell assembly as a mechanism for perceptual generalization.

摘要

For a species as oriented to objects as we are, recognizing objects is essential. Without this capacity, gathering and storing information would be useless, since any information that was stored could never be accessed or applied.; Recognizing objects is made difficult by the fact that the same object rarely if ever produces identical stimulus patterns. Every time an object is encountered, different features of the object are available in different combinations; further, there is variability both in these features and in the object's background. In order to manage this perpetual novelty, the ability to selectively generalize is essential. One must be able to generalize over the appropriate equivalence class of stimulus patterns, such that all of the stimulus patterns in an equivalence class indicate the presence of the same object. This ability is called perceptual generalization.; The focus of this dissertation is a simulation-based analysis of the perceptual generalization capabilities of Hebb's cell assembly. A cell assembly is a group of highly interconnected neurons that forms a reverberatory circuit capable of sustaining activity. Because of the reverberatory potential of this structure, nonidentical stimulus patterns can lead to the full activation of the same cell assembly. This provides a mechanism for producing the same response despite variability in the parts of the cell assembly that were initially activated.; Despite the importance of perceptual generalization and the potential that cell assembly theory has for providing a solution, the perceptual generalization capabilities of the cell assembly have never been explored via simulation. This dissertation: (1) explores these capabilities by studying the formation of cell assemblies through unsupervised perceptual learning. (2) demonstrates the utility of failure-driven experiments in incrementally developing a model of cell assembly functioning. (3) provides new insights on ways to visualize activity levels and connection strengths, two key factors in understanding the behavior of cell assemblies. (4) demonstrates the ability of cell assemblies to generalize despite the two most fundamental challenges to achieving perceptual generalization: variability and noise in the stimulus patterns.