A perceptual learning model based on topological representation
Mariofanna G. Milanova, Mark P. Wachowiak, Stuart H. Rubin, Adel Said Elmaghraby · 2002
This paper presents a model of perceptual learning based on an unsupervised self-organized algorithm for finding an efficient, low-dimensional representation of the data using natural image sequences. The current model is an extension of Barlow's redundancy reduction approach. An image sequence is a linear superposition of space-time functions convolved with a time varying coefficient signal to generate space-time inseparable basis functions. Independent component analysis (ICA) is one technique to determine these functions. In this work, it is proposed that dependencies not cancelled by ICA could define a topological order between the components. A learning algorithm, as well as its implementation in a cellular neural network that explicitly formalizes a topological order between the independent components, is presented.