Meta-learning for medical imaging: where are we after ICML?
Last week took place the renowned International Conference on Machine Learning (ICML) in Long Beach, CA. The conference is home to state-of-the-art research in Machine Learning, and while I could not get there myself, I try to always keep up with relevant applications for medical imaging.
This year I was particularly interested in the tutorial on meta-learning given by Chelsea Finn and Sergey Levine, leading researchers in the field. In short, meta-learning aims to learn efficiently from little amounts of data. It does so by learning on different tasks that are similar in structure. In radiology, where labeled data is scarce, this could become very useful. Currently meta-learning is in its infancy, and there are few applications where it is reliable enough to run in practice. However, the research direction is very interesting and it might be of great benefit to the medical domain. If we could manage to make algorithms learn from multiple medical tasks, we would greatly improve the effective amount of data to train medical applications.
But let’s take a step back. Even if the field improves tremendously, there is always the possibility that annotating more data for supervised learning will provide better outcomes than meta-learning. When talking about medical applications, we cannot morally (or legally) decide to trade off clinical outcomes for data availability. We will have to stay tuned for real-world applications before making the call.
Check the tutorial on meta-learning here. It contains links to the video, slides, and an introductory blog post on meta learning.