Events

Can Language Robustify Learning?

Machine learning is facing a robustness crisis. Despite reaching human-level performance on a wide range of benchmarks, state-of-the-art systems can be easily fooled by seemingly small perturbations that don’t affect humans. The problem is perhaps fundamental to learning: fitting huge low bias models can easily overfit the superficial statistics of the benchmarks. In this talk, we explore natural language as a way to address this problem by providing stronger supervision: instead of requiring the machine learning algorithm to infer a function from raw examples, we specify the function directly using language. We present two initial forays into this space: First, we convert natural language explanations to a function that labels unlabeled data, which can be used to train a predictor. Second, users interactively teach high-level concepts using natural language definitions.