AI paints a self-portrait
Image: IBM Research
Accompanying an NYT series on artificial intelligence today is a piece of art — seen above — that’s very unlike the newspaper's usual imagery: It represents AI, and it’s drawn by AI.
Why it matters: Artists are using increasingly powerful machine-learning algorithms to help produce fiction, film, and visual art. Incapable of creativity on their own, they can be programmed to act as a formidable artistic tool.
Developed by IBM Research, the algorithms that created this image were divided into three parts that together approximated a creative process.
- First, the system had to pick what to draw. It read about 3,000 NYT articles about AI and extracted the 30 most salient concepts, like robot, self-driving, and computing. It went on to unearth the 10 that were most representative, from which one — a human and robot shaking hands — was chosen.
- To build its own version, a generative adversarial network, or GAN, was trained on more than 1,000 existing images to create new ones.
- To match the newspaper’s style, a final step sampled past NYT imagery and applied the design to the AI-generated image.
Importantly, every step depended on human-generated content for training data.
- The algorithms synthesized thousands of artworks created previously by humans to make something novel.
- This means AI can help creative people make new things — but it can’t make something unique on its own, said John Smith, a fellow at IBM Research who worked on the project:
"Creativity itself, which is the leap of thought or imagination to create something completely new, different and valuable is still an essentially human ability."