Free Lunch for Few-shot Learning: Distribution Calibration

TL;DR

This paper introduces a distribution calibration method for few-shot learning that improves classification accuracy by transferring statistical information from well-sampled classes to under-sampled ones, without adding extra parameters.

Contribution

It proposes a novel distribution calibration technique that leverages Gaussian assumptions and transfer of class statistics to enhance few-shot learning performance.

Findings

Achieves ~5% accuracy improvement on miniImageNet.

Outperforms state-of-the-art methods without extra parameters.

Calibrated distributions accurately represent class features.

Abstract

Learning from a limited number of samples is challenging since the learned model can easily become overfitted based on the biased distribution formed by only a few training examples. In this paper, we calibrate the distribution of these few-sample classes by transferring statistics from the classes with sufficient examples, then an adequate number of examples can be sampled from the calibrated distribution to expand the inputs to the classifier. We assume every dimension in the feature representation follows a Gaussian distribution so that the mean and the variance of the distribution can borrow from that of similar classes whose statistics are better estimated with an adequate number of samples. Our method can be built on top of off-the-shelf pretrained feature extractors and classification models without extra parameters. We show that a simple logistic regression classifier trained using the features sampled from our calibrated distribution can outperform the state-of-the-art accuracy on two datasets (~5% improvement on miniImageNet compared to the next best). The visualization of these generated features demonstrates that our calibrated distribution is an accurate estimation.