Convolution pavée

Français

convolution pavée

Anglais

tiled convolution

Hidden layers within Convolutional Neural Networks reduce the number of parameters by "tying" together the adjacent N x N weights surrounding each input neuron. Each neuron in the hidden (convolutional) layer is only connected to an N x N grid of its surrounding neighbors (centered on a given neuron in the input layer), and the corresponding weights in each N x N grid connecting each hidden layer neuron to the input layer are the same (shared) across all hidden layer neurons. This weighted "local receptive field" is mathematically equivalent to a convolution operation (and a convolution is a special case of the more general "matrix multiplication" operation, expressed in fully-connected neural networks where the weights are "untied").

Source : quora Source : L'apprentissage profond, Ian Goodfellow, Yoshua Bengio et Aaron Courville Éd. Massot 2018