The wisdom of the crowd: ensemble learning and multi-modal imitation learning

The collective opion of a crowd is often considered to be better than that of a single expert. In this talk, we present some results that exploit this old idea in two machine learning problems: ensemble learning and multi-modal imitation learning.

Ensemble learning aims to find a collection of models with each having a different "opinion/view" on data, then combine them for prediction. We show that linear combinations can fail to generalize, while convex combinations can generalise for a broad class of Lipschitz-continuous loss functions. In addition, considering that there is little empirical study on learning a convex combination, we perform extensive experiments to study various greedy learning algorithms adapted for learning a convex combination. Our results suggest that a greedy algorithm based on a variant of the recently reinvogorated Frank-Wolfe algorithm achieves good performance without much hyperparameter tuning.

Multi-modal imitation learning aims to learn how to complete a task from demonstrations by possibly multiple experts. While some demonstration data consists of multiple behavior modes, typical approaches often learn a single behavior mode. We extend the maximum entropy inverse reinforcement learning framework to handle multi-modal data, and demonstrate its effectiveness on a multi-modal taxi-driving dataset.