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

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