The network learned measurements

Before AlexNet, a researcher decided in advance what to measure in a picture. Someone would write a procedure for finding edges, another for texture, another for corners, and the learning system only ever saw those measurements, never the picture itself. If a category needed something nobody had thought to measure, the system could not see it.

AlexNet received pixel values instead. Later studies visualized what activated units in related convolutional networks. They found examples of early units responding to edges and colours, while deeper units responded to more complex textures and object parts. This was an observed tendency in particular trained networks, not a fixed ladder that every network must follow.

This is the distinction from the last slide, in its most consequential form. The measurements had always been the designed part, written by researchers in advance. Now they were the learned part.

Everything around them stayed human: the images, the labels, the shape of the network, the training process, the test. What moved across the line was narrow, and it was enough to change the result.