Empirical Bayes Transductive Meta-Learning with Synthetic Gradients

Shell Xu Hu, Pablo Garcia Moreno, Yang Xiao, Xi Shen, Guillaume Obozinski, Neil D. Lawrence, Andreas Damianou
International Conference on Learning Representations, 2020.

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.

Cite this Paper

BibTeX
@InProceedings{Hu-empirical20, title = {Empirical {B}ayes Transductive Meta-Learning with Synthetic Gradients}, author = {Hu, Shell Xu and Garcia Moreno, Pablo and Xiao, Yang and Shen, Xi and Obozinski, Guillaume and Lawrence, Neil D. and Damianou, Andreas}, booktitle = {International Conference on Learning Representations}, year = {2020}, pdf = {https://openreview.net/attachment?id=Hkg-xgrYvH&name=original_pdf}, url = {http://inverseprobability.com/publications/empirical-bayes-transductive-meta-learning-with-synthetic-gradients.html}, 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. } }
Endnote
%0 Conference Paper %T Empirical Bayes Transductive Meta-Learning with Synthetic Gradients %A Shell Xu Hu %A Pablo Garcia Moreno %A Yang Xiao %A Xi Shen %A Guillaume Obozinski %A Neil D. Lawrence %A Andreas Damianou %B International Conference on Learning Representations %D 2020 %F Hu-empirical20 %U http://inverseprobability.com/publications/empirical-bayes-transductive-meta-learning-with-synthetic-gradients.html %X 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.
RIS
TY - CPAPER TI - Empirical Bayes Transductive Meta-Learning with Synthetic Gradients AU - Shell Xu Hu AU - Pablo Garcia Moreno AU - Yang Xiao AU - Xi Shen AU - Guillaume Obozinski AU - Neil D. Lawrence AU - Andreas Damianou BT - International Conference on Learning Representations DA - 2020/04/26 ID - Hu-empirical20 L1 - https://openreview.net/attachment?id=Hkg-xgrYvH&name=original_pdf UR - http://inverseprobability.com/publications/empirical-bayes-transductive-meta-learning-with-synthetic-gradients.html AB - 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. ER -
APA
Hu, S.X., Garcia Moreno, P., Xiao, Y., Shen, X., Obozinski, G., Lawrence, N.D. & Damianou, A.. (2020). Empirical Bayes Transductive Meta-Learning with Synthetic Gradients. International Conference on Learning Representations Available from http://inverseprobability.com/publications/empirical-bayes-transductive-meta-learning-with-synthetic-gradients.html.

Related Material