PRINCIPAL COMPONENT ANALYSIS LEARNING ALGORITHMS - A NEUROBIOLOGICAL ANALYSIS

The biological relevance of principal component analysis (PCA) learning algorithms is addressed by: (i) describing a plausible biological mechanism which accounts for the changes in synaptic efficacy implicit in Oja's 'Subspace' algorithm (Int. J. neural Syst. 1, 61 (1989)); and (ii) establishing a potential role for PCA-like mechanisms in the development of functional segregation. PCA learning algorithms comprise an associative Hebbian term and a decay term which interact to find the principal patterns of correlations in the inputs shared by a group of units. We propose that the presynaptic component of this decay could be regulated by retrograde signals that are translocated from the terminal arbors of presynaptic neurons to their cell bodies. This proposal is based on reported studies of structural plasticity in the nervous system. By using simulations we demonstrate that PCA-like mechanisms can eliminate afferent connections whose signals are unrelated to the prevalent pattern of afferent activity. This elimination may be instrumental in refining extrinsic cortico-cortical connections that underlie functional segregation.