2.23.21-AI-Isola

• Video details

  The last few years have seen an explosion of powerful generative models -- models that can synthesize fake faces, landscapes, text, audio and more. The results are fascinatingly realistic, but it's not immediately clear what they are useful for. We already have billions of images of faces, why do we need a model to make more? I will argue that the real power of these models is not their ability to make random fake data but that they make a new kind of data: data that comes bundled up with controllable latent variables. I will focus on deep generative models of images, which synthesize a photo given an input vector of latent variables. The latent variables are knobs that control what the output will look like: a user can tune them to change the lighting conditions in a photo, rotate objects, add or remove elements of a scene, and much more. I will show applications in image editing and scientific data visualization, and I will suggest that this new kind of data, sampled from deep generative models, can be thought of as data++: it looks just like regular data, but comes with extra functionality.

Locked Interactive transcript