- Tuesday 14 December 2021
- Online only
Jan van RIjn - LIACS
Title: Meta-learning: Making Sense out of Limited Data
Abstract: Modern AI solutions have a high dependency on large amounts of data. However, especially in life science applications, data is sometimes hard to acquire. The absence of this can lead to sub-optimal performance, which in turn limits the utility of the AI solution. In particular, deep neural networks can be very data-dependent. Data acquisition is an expensive part of the machine learning pipeline, which is in machine learning research projects often taken for granted.
However, for critical applications, access to labelled data can be the bottleneck. One way of addressing this data-dependency is by utilizing data from a similar source domain, and transferring the acquired knowledge to the target domain. The overarching field is called meta-learning.
In this talk, I will highlight several of the recent advances in meta-learning. I will briefly explain the assumptions that need to be met to apply it, and explain the basic methods MAML and Reptile. I will further explain some work that my group at LIACS does to better understand the underlying techniques. I will showcase some of the use-cases in life-sciences where we are successfully applied meta-learning to improve the utility of few data points.