EMO: Episodic Memory Optimization for Few-Shot Meta-Learning

Yingjun Du, Jiayi Shen, Xiantong Zhen, Cees G. M. Snoek

Keywords: Episodic memory, Meta-learning, Few-shot learning, Optimization

TL;DR: We propose an episodic memory optimization for meta-learning, which we call EMO, that retains the gradient history of past experienced tasks in external memory.

Abstract: For few-shot meta-learning, gradient descent optimization is challenging due to the limited number of training samples per task. Inspired by the human ability to recall past learning experiences from the brain’s memory, we propose an episodic memory optimization for meta-learning, which we call EMO, that retains the gradient history of past experienced tasks in external memory. It enables few-shot learning in a memory-augmented way by leveraging the meta-learning setting and learns to retain and recall the learning process of past training tasks for gradient descent optimization. By doing so, EMO nudges the parameter updates in the right direction, even when the gradients provided by a limited number of examples are uninformative. Additionally, we prove theoretically that our algorithm converges for smooth, strongly convex objectives. EMO is generic, flexible, and model agnostic, making it a simple plug-and-play optimizer seamlessly embedded into existing optimization-based meta-learning approaches. Empirically, EMO scales well with most of the few-shot classification benchmarks, and our experiments show that the optimization-based meta-learning method enjoys accelerated convergence and improved performance with EMO.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: General Machine Learning (ie none of the above)