Learning from Natural Language Supervision
Abstract
Humans can efficiently learn and communicate new knowledge about the world through natural language (e.g, the concept of important emails may be described through explanations like ‘late night emails from my advisor are definitely important’). Can machines be similarly taught new tasks and behavior through natural language interactions with their users? In this talk, we'll explore three approaches towards enabling language-based learning in context of classifications tasks. In the first part of the talk, we'll consider how language can be leveraged for interactive feature space construction for machine learning tasks. I'll present a method for jointly (1) learning to interpret open-ended natural language explanations, and (2) learning the classification models, by using explanations in conjunction with a small number of labeled examples of the concept. Secondly, we'll examine an approach for using language as a substitute for labeled data for generalization of machine learning models, which leverages the semantics of quantifiers in everyday language (`definitely', `sometimes', etc.) to enable machine learning with limited or no labeled data. Finally, we'll briefly explore how conversational agents can accelerate the learning process through student-teacher dialog.