5.5.22-Efficient-AI-Antonio-Torralba

• Video details

  The importance of data in modern computer vision is hard to overstate. The ImageNet dataset, with its millions of labelled images, is widely thought to have spurred the era of deep learning, and since then the scale of vision datasets has been increasing at a rapid pace. These datasets come with costs: curation is expensive, and they inherit human biases. To counter these costs, interest has surged in learning with unlabeled images as it avoids the curation efforts, or using simulated environments, but content creation is also labor intensive. In our work we go a step further and ask if we can do away with real image datasets entirely, instead learning from noise processes. Noise processes produce images that are reminiscent of abstract art, where images contain textures and shapes, but there are no recognizable objects. Our findings show that good performance on real images can be achieved even with training images that are far from realistic.

Locked Interactive transcript