OpenAI's iGPT algorithm in action. Credit: OpenAI
An AI algorithm is capable of automatically generating realistic-looking images from bits of pixels.
Why it matters: The achievement is the latest evidence that AI is increasingly able to learn from and copy the real world in ways that may eventually allow algorithms to create fictional images that are indistinguishable from reality.
What's new: In a paper presented at this week's International Conference on Machine Learning, researchers from OpenAI showed they could train the organization's GPT-2 algorithm on images.
- GPT-2 is best known as a text-generating algorithm, capable of absorbing the structure of language by training on billions of words off the internet and then "writing" passable text from a simple prompt.
- After training on images, the OpenAI algorithm — now called iGPT — could be fed a visual prompt in the form of half an image and then fill in the rest of it.
How it works: Unlike algorithms that use supervised learning, which requires laborious amounts of labeled data, iGPT and GPT-2 use unsupervised learning on unlabeled data. That means much less human effort.
- As Karen Hao points out in MIT Tech Review, the fact that both iGPT and GPT-2 use the same algorithm for different purposes is early evidence of more generalizable machine intelligence — which is, after all, OpenAI's aim.
The bottom line: A world where you can ask a robot to write, or create an image of, anything may be just around the corner, which is both exciting and terrifying.
Go deeper: Rooting out AI bias