Strength from Weakness: Fast Learning Using Weak Supervision

TL;DR

This paper demonstrates that weak supervision can significantly accelerate learning rates for strongly labeled tasks, achieving faster convergence even when strong labels alone would be slow, with theoretical and empirical support.

Contribution

The paper provides a theoretical analysis showing how weak labels can improve learning rates for strong tasks and validates this with empirical experiments.

Findings

Weak labels can accelerate learning to a fast rate of O(1/n).

Acceleration depends on the number of weak labels and task relation.

Empirical results confirm theoretical predictions.

Abstract

We study generalization properties of weakly supervised learning. That is, learning where only a few "strong" labels (the actual target of our prediction) are present but many more "weak" labels are available. In particular, we show that having access to weak labels can significantly accelerate the learning rate for the strong task to the fast rate of $\mathcal{O}(\nicefrac1n)$, where $n$ denotes the number of strongly labeled data points. This acceleration can happen even if by itself the strongly labeled data admits only the slower $\mathcal{O}(\nicefrac{1}{\sqrt{n}})$ rate. The actual acceleration depends continuously on the number of weak labels available, and on the relation between the two tasks. Our theoretical results are reflected empirically across a range of tasks and illustrate how weak labels speed up learning on the strong task.