Empirical Bayes Transductive Meta-Learning with Synthetic Gradients

TL;DR

This paper introduces a transductive meta-learning method that uses synthetic gradients and an empirical Bayes framework to improve few-shot classification, achieving state-of-the-art results on Mini-ImageNet and CIFAR-FS.

Contribution

It develops a novel amortized variational inference approach coupling variational posteriors with a meta-model using synthetic gradients, advancing meta-learning techniques.

Findings

Outperforms previous state-of-the-art on Mini-ImageNet and CIFAR-FS benchmarks.

Demonstrates effectiveness of synthetic gradients in zero-shot learning tasks.

Provides a new framework for transductive meta-learning with empirical Bayes.

Abstract

We propose a meta-learning approach that learns from multiple tasks in a transductive setting, by leveraging the unlabeled query set in addition to the support set to generate a more powerful model for each task. To develop our framework, we revisit the empirical Bayes formulation for multi-task learning. The evidence lower bound of the marginal log-likelihood of empirical Bayes decomposes as a sum of local KL divergences between the variational posterior and the true posterior on the query set of each task. We derive a novel amortized variational inference that couples all the variational posteriors via a meta-model, which consists of a synthetic gradient network and an initialization network. Each variational posterior is derived from synthetic gradient descent to approximate the true posterior on the query set, although where we do not have access to the true gradient. Our results on the Mini-ImageNet and CIFAR-FS benchmarks for episodic few-shot classification outperform previous state-of-the-art methods. Besides, we conduct two zero-shot learning experiments to further explore the potential of the synthetic gradient.