Deep Learning with PyTorch Quick Start Guide

Convolutional networks

So far, we have used fully connected layers in our networks, where each input unit represents a pixel in an image. With convolutional networks, on the other hand, each input unit is assigned a small localized receptive field. The idea of the receptive field, like ANNs themselves, is modelled on the human brain. In 1958, it was discovered that neurons in the visual cortex of the brain respond to stimuli in a limited region of a field of vision. More intriguing is that sets of neurons respond exclusively to certain basic shapes. For example, a set of neurons may respond to horizontal lines, while others respond only to lines at other orientations. It was observed that sets of neurons could have the same receptive field, ...