Chapter 6. Convolutional Neural Networks
Convolutional neural networks allow deep networks to learn functions on structured spatial data such as images, video, and text. Mathematically, convolutional networks provide tools for exploiting the local structure of data effectively. Images satisfy certain natural statistical properties. Let’s assume we represent an image as a two-dimensional grid of pixels. Parts of an image that are close to one other in the pixel grid are likely to vary together (for example, all pixels corresponding to a table in the image are probably brown). Convolutional networks learn to exploit this natural covariance structure in order to learn effectively.
Convolutional networks are a relatively old invention. Versions of convolutional networks have been proposed in the literature dating back to the 1980s. While the designs of these older convolutional networks were often quite sound, they required resources that exceeded hardware available at the time. As a result, convolutional networks languished in relative obscurity in the research literature.
This trend reversed dramatically following the 2012 ILSVRC challenge for object detection in images, where the convolutional AlexNet achieved error rates half that of its nearest competitors. AlexNet was able to use GPUs to train old convolutional architectures on dramatically larger datasets. This combination of old architectures with new hardware allowed AlexNet to dramatically outperform the state of the art ...