ESP Biography

How do machine learning models produce images of dogs? A large number of random building blocks, when put together, can produce patterns with a great deal of structure: individual blocks are unpredictable, but collective behavior is lawful. This is a theme underlying huge swathes of mathematics, physics, and machine learning. We'll drive this idea home by first studying "diffusion" --- the motion of an object that stumbles around randomly --- via an in-class collective experiment. Then, we’ll switch gears to machine learning. You’ll see how diffusion models --- a cutting-edge technique often used in AI image generation --- rely on the same mathematical ideas. With hands-on visualizations, you’ll witness how an image of a dog is gradually destroyed with noise, and a new dog is reconstructed! From order to randomness, and back again.