Competitive Learning

A Competitive learning is an artificial neural network learning process where different neurons or processing elements compete on who is allowed to learn to represent the current input. In its purest form competitive learning is in the so-called winner-take-all networks where only the neuron that best represents the input is allowed to learn. Since all neurons learn to better represent the kinds of inputs they already are good at representing, they become specialized to represent different kinds of inputs. For vector-valued inputs and representations, the input becomes quantized to the unit having the closest representation (model), and the representations are adapted to minimize the representation error using stochastic gradient descent.

Competitive learning networks have been studied as models of how receptive fields and feature detectors, such as orientation-selective visual neurons, develop in neural networks. The same process is at work in online K-means clustering, and variants...