Transductive inference is widely used in few-shot learning, as it leverages the statistics of the unlabeled query set of a few-shot task, typically yielding substantially better performances than its inductive counterpart. The current few-shot benchmarks use perfectly class-balanced tasks at inference. We argue that such an artificial regularity is unrealistic, as it assumes that the marginal label probability of the testing samples is known and fixed to the uniform distribution. In fact, in realistic scenarios, the unlabeled query sets come with arbitrary and unknown label marginals. We introduce and study the effect of arbitrary class distributions within the query sets of few-shot tasks at inference, removing the class-balance artefact. Specifically, we model the marginal probabilities of the classes as Dirichlet-distributed random variables, which yields a principled and realistic sampling within the simplex.

Transductive inference is widely used in few-shot learning, as it leverages the statistics of the unlabeled query set of a few-shot task, typically yielding substantially better performances than its inductive counterpart. The current few-shot benchmarks use perfectly class-balanced tasks at inference. We argue that such an artificial regularity is unrealistic, as it assumes that the marginal label probability of the testing samples is known and fixed to the uniform distribution. In fact, in realistic scenarios, the unlabeled query sets come with arbitrary and unknown label marginals. We introduce and study the effect of arbitrary class distributions within the query sets of few-shot tasks at inference, removing the class-balance artefact.

We propose a novel transductive inference framework for metric-based meta-learning models, which updates the prototype of each class with the confidence-weighted average of all the support and query samples. However, a caveat here is that the model confidence may be unreliable, which could lead to incorrect prediction in the transductive setting. To tackle this issue, we further propose to meta-learn to assign correct confidence scores to unlabeled queries. Specifically, we meta-learn the parameters of the distance-metric, such that the model can improve its transductive inference performance on unseen tasks with the generated confidence scores. We also consider various types of uncertainties to further enhance the reliability of the meta-learned confidence. We combine our transductive meta-learning scheme, Meta-Confidence Transduction (MCT) with a novel dense classifier, Dense Feature Matching Network (DFMN), which performs both instance-level and feature-level classification without global average pooling and validate it on four benchmark datasets. Our model achieves state-of-the-art results on all datasets, outperforming existing state-of-the-art models by 11.11% and 7.68% on miniImageNet and tieredImageNet dataset respectively. Further qualitative analysis confirms that this impressive performance gain is indeed due to its ability to assign high confidence to instances with the correct labels.