Abstract

We develop a transductive meta-learning method that uses unlabelled instancesto improve few-shot image classification performance. Our approach combines aregularized Mahalanobis-distance-based soft k-means clustering procedure with amodified state of the art neural adaptive feature extractor to achieve improvedtest-time classification accuracy using unlabelled data. We evaluate our methodon transductive few-shot learning tasks, in which the goal is to jointlypredict labels for query (test) examples given a set of support (training)examples. We achieve state-of-the-art performance on the Meta-Dataset,mini-ImageNet and tiered-ImageNet benchmarks.

 

Quick Read (beta)