Example of a DeepDream artwork. The approach makes use of an artificial neural network that can seek and enhance patterns in images, with startling results.

2015

COMPUTER ART AND DEEPDREAM

According to essayist Jonathan Swift (1667–1745), “Vision is the art of seeing things invisible.” This idea of seeing new patterns at the edges of art, science, and mathematics certainly applies to many categories of art produced with the aid of computers, algorithms, neural networks, and other forms of AI. Early explorations of com-puter art include the work of Desmond Paul Henry (1921–2004) and his bombsight analog computer-based drawing machines, starting around 1961. American engineer A. Michael Noll (b. 1939) became famous in 1962 for exploring random and algorithmic processes to produce visual art, and British-born artist Harold Cohen (1928–2016) created AARON—an AI computer-drawing program that could autonomously produce art—in 1968.

A more recent example of computer art includes the collaboration of many users with DeepDream, a 2015 computer-vision program created by Google engineer Alexander Mordvintsev and colleagues. The approach makes use of an artificial neural network (ANN) that can seek and enhance patterns in images—with startling results. To better understand DeepDream, consider that neural nets may be trained to classify and recognize features in an input image (e.g., chipmunks or stop signs) based on numerous “training” images. By running the neural network “in reverse,” DeepDream looks for patterns in images and amplifies them in a manner somewhat reminiscent of when we gaze up at the clouds and begin to see animal-like shapes. For artificial neural nets, each layer progressively extracts higher-level features—for example, the first layer might be sensitive to corners and edges, while layers closer to the output neurons may be examining complex features. Not only are the resulting pictures fascinating to study and brimming with detail, but they can provide a sense of the level of abstraction on which a particular ANN layer is working.

The resemblance of DeepDream artwork to hallucinations experienced by users of certain mind-altering drugs suggests that DeepDream might help researchers better understand how artificial neural networks relate to actual neural networks in the brain’s visual cortex. Furthermore, the program could help illuminate how the brain attempts to find pattern and meaning.