“One of the most powerful things about this technology is that, like the DALL-E, it does what you ask it to do,” said Nate Bennett, one of the researchers in the UW lab. “From one cue, it can generate countless designs.”
The rise of OpenAI
The San Francisco-based company is one of the most ambitious artificial intelligence labs in the world. Here are some recent developments.
To generate the images, DALL-E relies on what AI researchers call a neural network, a mathematical system that loosely mimics the brain’s network of neurons. The technology recognizes commands you give to your smartphone, enables self-driving cars to recognize (and avoid) pedestrians, and translates languages on services like Skype.
Neural networks learn skills by analyzing large amounts of numerical data. For example, it can learn to recognize corgis by pinpointing patterns in thousands of photos of corgis. With DALL-E, the researchers built a neural network that looks for patterns as it analyzes millions of digital images and text captions describing what each image depicts. In this way, it learned to recognize connections between images and words.
When you describe an image to DALL-E, the neural network generates a set of key features that the image is likely to contain. One feature might be the curve of the teddy bear’s ears. Another possibility is a line on the edge of the skateboard. Then, a second neural network — called a diffusion model — generates the pixels needed to achieve those features.
Diffusion models are trained on a series of images in which noise — imperfections — is gradually added to a photo until it becomes a sea of random pixels. When it analyzes these images, the model learns to run the process in reverse. When you give it random pixels, it removes noise, turning those pixels into a coherent image.
At Washington University, other academic labs and new start-ups, researchers are using similar techniques in an effort to create new proteins.
Proteins start out as strings of chemical compounds that twist and fold into three-dimensional shapes that define how they behave. In recent years, AI labs like DeepMind, owned by Google parent company Alphabet, have shown that neural networks can accurately guess the three-dimensional shape of any protein in the body, based solely on the smaller compounds it contains—a huge scientific advance.