Plenary Talk: Image Models and Unsupervised Learning

The importance of large databases of images in modern computer vision is hard to overstate. In the last decade, the scale of vision datasets has been increasing at a rapid pace, but these datasets come with costs: image collection and data annotation are expensive. To counter these costs, interest has surged in unsupervised learning, as it avoids the curation efforts. But images are still hard to collect. An alternative to real images is to use simulations with graphics engines, but content creation is also costly. In this talk, we will go a step further and ask if we can do away with real image datasets—or simulations—entirely, and instead learn from simple generative processes inspired by early work on the statistics of natural images. I will give a tour through the history of classical models of natural images, and show the power of simple generative image models in training visual representations that can rival those learned from real images. I will describe several generative models that produce images reminiscent of abstract art—images with textures and shapes, but no recognizable objects. Surprisingly, we will show that good performance on real images can be achieved even when the training images are far from realistic.