Online spectral clustering and the neural mechanisms of concept formation

Abstract

Spectral clustering can provide surprising performances. As all kernel methods, is uses a similarity matrix, whose size grows with re2, and it requires to solve a possibly large eigenproblem. In this paper we focus on a method for spectral embedding of stream data, modeled as an unbounded quantity of input observation. A second purpose of this work is to analyze the proposed method and compare it with traditional neural network implementations: current knowledge about computations in neurons and the brain does not contrast with the computing primitives required for a local implementation of the proposed technique. A hypothesis stemming from this work could be that concept formation and discrimination in neurons and the brain could be explained by a spectral embedding framework.