Adaptive Task Sampling and Variance Reduction for Gradient-Based Meta-Learning


Zhuoqun Liu (Shanghai Jiao Tong University),* Yuankun Jiang (Shanghai Jiao Tong University), Chenglin Li (Shanghai Jiao Tong University), Wenrui Dai (Shanghai Jiao Tong University), Junni Zou (Shanghai Jiao Tong University), Hongkai Xiong (Shanghai Jiao Tong University)
The 33rd British Machine Vision Conference

Abstract

Meta-learning enables fast adaptation of the trained model to new tasks by exploiting the similarity between tasks sampled for training, leading to its success in few-shot learning and domain adaptation. Conventional meta-learning paradigm, however, treats different tasks equally important and thus samples them at a uniform distribution for training, which may result in a sub-optimal performance with high variance introduced to the gradients. To address this, in this paper, we develop a novel adaptive task sampling and variance reduction (ATSVR) method for gradient-based meta-learning. Built upon gradient-based meta-learning framework, we are able to assign different importance weights to the training tasks, by leveraging the importance sampling technique to approximately manipulate the sampling distribution according to a target distribution that is updated iteratively towards minimising the meta-objective. In addition, update of this target distribution is also enforced to reduce the variance of gradient estimate at each iteration. Empirical evaluations on a regression task demonstrate the performance gain by introducing adaptive task sampling to meta-learning, while those on the few-shot learning task on two benchmarks show that our ATSVR outperforms state-of-the-art adaptive sampling-based baselines, such as meta-learning with adaptive task scheduler.

Video



Citation

@inproceedings{Liu_2022_BMVC,
author    = {Zhuoqun Liu and Yuankun Jiang and Chenglin Li and Wenrui Dai and Junni Zou and Hongkai Xiong},
title     = {Adaptive Task Sampling and Variance Reduction for Gradient-Based Meta-Learning},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year      = {2022},
url       = {https://bmvc2022.mpi-inf.mpg.de/0876.pdf}
}


Copyright © 2022 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection