Contact: Arnout Devos
keywords: deep learning, meta-learning, few-shot learning
In few-shot classification, we are interested in learning algorithms that train a classifier from only a handful of labeled examples. Meta-learning is a successful approach for few-shot learning. A more natural extension of the few-shot learning problem is a setting where, next to the few labeled examples (shots), there are many unlabeled examples present. A real-life application would be a camera mounted on a car, where only few images of all the collected images need to be labeled to build a performant classifier. The idea of the project is to investigate state-of-the-art meta-learning algorithms in the setting of semi-supervised learning.
The first part of the project will consist mostly of a literature search on few-shot learning algorithms and their extensions to semi-supervised learning. After investigating the benefits and drawbacks of each method, we will implement some methods and experimentally compare their performance for the case of few-shot semi-supervised learning.
This semester project is targeted towards a master's student.
Meta-Learning for Semi-Supervised Few-Shot Classification [ICLR 2018]