2

I was reading the AlexNet paper (namely "ImageNet Classification with Deep Convolutional Neural Networks"). The Introduction section states "stationarity of statistics" and "locality of pixel dependencies" as the two strong assumptions made in convolutional neural networks. I understand that the latter refers to the existence of local patterns in the image, but what is the meaning of "stationarity of statistics"?

Vaibhav Jain
  • 195
  • 1
  • 10

1 Answers1

0

Stationarity usually refers to properties which do not change over time, but in the context of CNNs it refers to how separate neighbourhoods of pixels tend to display similar patterns. This is why the same set of filters can be used to capture local correlations at different locations in the input image, and at different locations across multiple input images.

Benjamin Rodrigues de Miranda
  • 23
  • 1
  • 8