Unlocking the Full Potential of Small Data with Diverse Supervision

TL;DR

This paper addresses the challenge of small data scenarios in deep learning by leveraging diverse supervision from related tasks, demonstrating that dense labeling of limited images can significantly improve model performance.

Contribution

It introduces the LRDS benchmark using ADE20K for diverse supervision in small data settings and explores novel scenarios like Scarce-Class and Scarce-Image to enhance generalization.

Findings

Densely labeling few images improves performance in small data settings.

Using diverse supervision sources enhances model generalization.

The proposed benchmarks reveal the benefits of rich annotations for small data tasks.

Abstract

Virtually all of deep learning literature relies on the assumption of large amounts of available training data. Indeed, even the majority of few-shot learning methods rely on a large set of "base classes" for pretraining. This assumption, however, does not always hold. For some tasks, annotating a large number of classes can be infeasible, and even collecting the images themselves can be a challenge in some scenarios. In this paper, we study this problem and call it "Small Data" setting, in contrast to "Big Data". To unlock the full potential of small data, we propose to augment the models with annotations for other related tasks, thus increasing their generalization abilities. In particular, we use the richly annotated scene parsing dataset ADE20K to construct our realistic Long-tail Recognition with Diverse Supervision (LRDS) benchmark by splitting the object categories into head and tail based on their distribution. Following the standard few-shot learning protocol, we use the head classes for representation learning and the tail classes for evaluation. Moreover, we further subsample the head categories and images to generate two novel settings which we call "Scarce-Class" and "Scarce-Image", respectively corresponding to the shortage of samples for rare classes and training images. Finally, we analyze the effect of applying various additional supervision sources under the proposed settings. Our experiments demonstrate that densely labeling a small set of images can indeed largely remedy the small data constraints.