A mechanism for the development of feature detecting neurons

The mammalian retina and visual cortex contain feature detecting neurons with a remarkable similarity to neurons in artificial neural networks for principal component analysis. Hebbian-type learning is one of the mechanisms responsible for the development of such neurons. It does not model, however, control of the number of neurons that develop in response to input. We propose a mechanism that adaptively controls this number. The mechanism utilizes the variances of neurons' outputs and encodes them as the lengths of the neural network's synaptic weight vectors, thus allowing only the synapses of those neurons to develop that represent significant information about the neural network's input and suppressing neurons' synapses that don't.