The role of constraints in Hebbian learning
Neural Computation
The role of weight normalization in competitive learning
Neural Computation
An analysis of synaptic normalization in a general class of Hebbian models
Neural Computation
| Hi-index | 0.01 |
In the mammalian brain, inputs from the two eyes compete with each other during development for synaptic contacts with postsynaptic targets. Traditionally, binocular competition has been modeled using a Hebbian-like plasticity rule together with a fixed-sum constraint on total connectivity. Previous work has shown the certain, but not all, formulations of the constrained-connectivity model lead to binocular competition. These differences between model formulations had been analyzed using a correlation-based approach, particularly relevant for the developmental period after eye opening, when the two eyes are normally active together but at varying levels of inter-ocular correlation. More recently, experimental work has established a role for Hebbian plasticity in binocular competition also prior to eye opening. In contrast to later developmental stages, at this pre-vision stage, activity of the two eyes is largely de-correlated. The goal of the present study is twofold. First, a simplified analysis of the constrained-competition model is proposed for the case of de-correlated binocular activity as observed prior to eye opening. This analysis provides a more intuitive explanation for the sensitivity of the constrained-connectivity model to formulation details. Second, based on recent experimental findings, this work proposes a new, modified model in which binocular competition increases the total connectivity instead of operating under a fixed-sum constraint. Simulations and analysis show that this form of ''growth-promoting'' competition is less sensitive to formulation details and may therefore be biologically advantageous at early stages of visual development.