Timezone: »

 
Poster
Memory Efficient Meta-Learning with Large Images
John Bronskill · Daniela Massiceti · Massimiliano Patacchiola · Katja Hofmann · Sebastian Nowozin · Richard Turner

Tue Dec 07 08:30 AM -- 10:00 AM (PST) @ Virtual
in Poster Session 1 »

Meta learning approaches to few-shot classification are computationally efficient at test time, requiring just a few optimization steps or single forward pass to learn a new task, but they remain highly memory-intensive to train. This limitation arises because a task's entire support set, which can contain up to 1000 images, must be processed before an optimization step can be taken. Harnessing the performance gains offered by large images thus requires either parallelizing the meta-learner across multiple GPUs, which may not be available, or trade-offs between task and image size when memory constraints apply. We improve on both options by proposing LITE, a general and memory efficient episodic training scheme that enables meta-training on large tasks composed of large images on a single GPU. We achieve this by observing that the gradients for a task can be decomposed into a sum of gradients over the task's training images. This enables us to perform a forward pass on a task's entire training set but realize significant memory savings by back-propagating only a random subset of these images which we show is an unbiased approximation of the full gradient. We use LITE to train meta-learners and demonstrate new state-of-the-art accuracy on the real-world ORBIT benchmark and 3 of the 4 parts of the challenging VTAB+MD benchmark relative to leading meta-learners. LITE also enables meta-learners to be competitive with transfer learning approaches but at a fraction of the test-time computational cost, thus serving as a counterpoint to the recent narrative that transfer learning is all you need for few-shot classification.

Keywords: Optimization · Machine Learning · Vision · Meta Learning · Transfer Learning · Few Shot Learning

Author Information

John Bronskill (University of Cambridge)
Daniela Massiceti (Microsoft Research)
Massimiliano Patacchiola (University of Cambridge)

Massimiliano is a postdoctoral researcher at the University of Cambridge in the Machine Learning Group. He is interested in efficient learning (few-shot, self-supervised, meta-learning), Bayesian methods (Gaussian processes), and reinforcement learning. Previously he has been a postdoctoral researcher at the University of Edinburgh and an intern in the Camera Platform team at Snapchat.

Katja Hofmann (Microsoft Research)

Dr. Katja Hofmann is a Principal Researcher at the [Game Intelligence](http://aka.ms/gameintelligence/) group at [Microsoft Research Cambridge, UK](https://www.microsoft.com/en-us/research/lab/microsoft-research-cambridge/). There, she leads a research team that focuses on reinforcement learning with applications in modern video games. She and her team strongly believe that modern video games will drive a transformation of how we interact with AI technology. One of the projects developed by her team is [Project Malmo](https://www.microsoft.com/en-us/research/project/project-malmo/), which uses the popular game Minecraft as an experimentation platform for developing intelligent technology. Katja's long-term goal is to develop AI systems that learn to collaborate with people, to empower their users and help solve complex real-world problems. Before joining Microsoft Research, Katja completed her PhD in Computer Science as part of the [ILPS](https://ilps.science.uva.nl/) group at the [University of Amsterdam](https://www.uva.nl/en). She worked with Maarten de Rijke and Shimon Whiteson on interactive machine learning algorithms for search engines.

Sebastian Nowozin (Microsoft Research Cambridge)
Richard Turner (University of Cambridge)

More from the Same Authors