Unsupervised Embedding Adaptation via Early-Stage Feature Reconstruction for Few-Shot Classification

TL;DR

This paper introduces ESFR, an unsupervised embedding adaptation method that enhances few-shot classification by focusing on early-stage features, leading to state-of-the-art results across multiple datasets.

Contribution

The paper presents a novel unsupervised adaptation scheme, ESFR, that leverages early-stage feature reconstruction and early stopping to improve few-shot learning performance.

Findings

ESFR improves baseline methods across all standard settings.

Combining ESFR with transductive methods achieves state-of-the-art results.

ESFR yields 1.2% to 2.0% accuracy improvements in 1-shot classification.

Abstract

We propose unsupervised embedding adaptation for the downstream few-shot classification task. Based on findings that deep neural networks learn to generalize before memorizing, we develop Early-Stage Feature Reconstruction (ESFR) -- a novel adaptation scheme with feature reconstruction and dimensionality-driven early stopping that finds generalizable features. Incorporating ESFR consistently improves the performance of baseline methods on all standard settings, including the recently proposed transductive method. ESFR used in conjunction with the transductive method further achieves state-of-the-art performance on mini-ImageNet, tiered-ImageNet, and CUB; especially with 1.2%~2.0% improvements in accuracy over the previous best performing method on 1-shot setting.