Data-efficient Machine Learning with People and Robots

In this talk I will provide an introduction to several data-efficient machine learning approaches. The focus will be on algorithms that actively help to reduce the number of expensive data samples. I will describe Bayesian Optimization - a global search method that determines the optimal next "query". It achieves a balance between sampling points in the promising regions and exploring the uncertain parts of the space. The talk will include examples of using Bayesian Optimization in robotics. I will also discuss the challenge and opportunity of working with noisy data. High levels of noise are particularly frequent when data comes from human studies. The challenge is to make sure that machine learning algorithms employed are robust enough. The opportunity is to discover an interesting and complex signal despite the noise. The final part of the talk will include most recent developments in active learning: combination with deep networks. This direction could be useful for cases with high-dimensional samples (video/speech/sensor data).