Sliding the same detector

A vision network needs a way to reuse what it learns across different parts of an image.

A photograph is large. A modest one has hundreds of thousands of pixels. Connecting every pixel to every neuron in the next layer would need many separate weights.

Worse, the effort would be wasted. A network that learned to spot an ear in the top-left corner would know nothing about ears in the middle, because those are different pixels feeding different weights. Move the cat and it starts again.

The fix is to stop looking at the whole picture at once.

Take a small detector, a window a few pixels across, and slide it over the image. Across and down, until it has visited everywhere. The same detector, with the same weights, at every position.

Now the same detector can respond to a similar local pattern in many positions. This property is called translation equivariance: move the pattern, and the detector's response moves with it. The whole network is not automatically indifferent to position, but it no longer needs a separate detector for every location.

This is called a , and a network built from these is a convolutional neural network. AlexNet was one.