Café 微積分

Most ANNs are organized into computational layers. Whether those layers have memory, feedback, gating or other sophisticated features, it still often follows that ANNs are organized in some hierarchy. Each layer of the hierarchy will take input from the previous layer, process it, and pass it forward to the next layer. The representation that the ANN produces is abstracted from its original form, but somehow it remains related to its original. The machine, like humanity, will abstract reality to isolate only the details of interest to the abstracter.

With each layer of processing, the information fed into the ANN becomes more abstract.

At first, the input might be something meaningful to humans, like an image.

The first layer examines the input to detect useful patterns: in images, these may be things like tones, edges, hue, but also less interpret-able things like "the average amount of yellow in pixels (32,91), (2,47) and (86,204).

The next layer takes the first layers' results and processes them: the amount of yellow in the top left combined with the number of vertical edges there. I don't know what that's good for, but there may be a neuron that cares about this anyways.

After several layers of this, the output of the ANN is typically something again meaningful to humans, such as a DeepFake or a label that tells you if the image had a dog or cat.

Jean-Michel Basquiat, Warrior (1982)

With each layer of processing, the information fed into the ANN becomes more abstract.

At first, the input might be something meaningful to humans, like an image.

The first layer examines the input to detect useful patterns: in images, these may be things like tones, edges, hue, but also less interpret-able things like "the average amount of yellow in pixels (32,91), (2,47) and (86,204)."

The next layer takes the first layers' results and processes them: the amount of yellow in the top left combined with the number of vertical edges there. I don't know what that's good for, but there may be a neuron that cares about this anyways.

After several layers of this, the output of the ANN is typically something again meaningful to humans, such as a DeepFake or a classification of an image as "scenery" or "portrait".

While virtually meaningless to humans, the abstract representations produced by the network are extremely useful. The way in which Artbreeder allows you manipulate a slider which transforms an image is extremely cool, but this is only possible because the ANN knows how to convert between images and some low-dimensional representation of them (it's not really low dimensions like 1,2 and 3, but compared to the 3,000,000 coordinates needed to describe a 100 by 100 pixel image, we can be generous with what we call low dimensional).