Fri, 19 May 2017
CNNs are characterized by their use of a group of neurons typically referred to as a filter or kernel. In image recognition, this kernel is repeated over the entire image. In this way, CNNs may achieve the property of translational invariance - once trained to recognize certain things, changing the position of that thing in an image should not disrupt the CNN's ability to recognize it. In this episode, we discuss a few high-level details of this important architecture.