A type of neural network that learns features via convolutional kernels.
In images, pixels that are spatially close are often related to each other — this assumption can be used as an inductive bias in neural networks. Perceptrons in the next layers can be connected only to locally close perceptrons of the previous layers.
Further, constructing the weight matrix using a specific repeating weight structure can make the matrix multiplication a convolution, which can be studied as filters for images making the models better for interpretability — at least for the initials perceptron layers.
This page is a part of Machine Learning.