Learning without Annotation

Supervised learning has worked very well in practice but yet they require exhaustive labelled data, posing a barrier to widespread adoption. My research works towards a holy grail of unsupervised learning that can either learn without annotation completely or with noisy and easy to obtain labels. I like to think about "representation learning" algorithms that learn to capture the inherent structure of the data itself (e.g. the distribution or invariances in properties), or multimodality, where an assortment of images, text, videos, audio clips, etc each provides a stronger learning signal we could use to build a richer and more abstract representation.