convolutional neural network

A convolutional neural network performs the same transformation to many small patches during image processing or to consecutive tokens in time series or sequential data. This may be achieved by having lots of different nodes for each patch, but where the weights are 'clamped' to be the same, or by reusing a small network. In both cases backpropagation or other learning rules have to be modified. The technique derives its name from mathematical convolutions used in linear time-series analysis.