Computer Sciences Colloquium - Discovering Analogies Between Domains

This talk will describe my past and ongoing work on discovering analogies between domains without supervision. Although Humans often do not require supervision to make analogies in completely new domains, this is still difficult for machines. Recently great progress was made by using adversarial training – a powerful yet tricky method. Although adversarial methods have had great success, they have significant failings such as difficult saddle point training and mode collapse. The drawbacks of generative adversarial models significantly limit their breadth of applicability, motivating research into alternative non-adversarial methods. I will describe my research on novel non-adversarial methods for unsupervised matching and mapping in the text and image domains. It will be demonstrated that unconditional generative image models are an important step towards making unsupervised analogies. I will then describe my most recent work on non-adversarial unconditional generative models, paving the way to truly non-adversarial analogies.