--- UPDATE ---: This project is now taken by Tim van Engeland
Meta-learning (also referred to as learning to learn) is a set of Machine Learning techniques that aim to learn quickly from a few given examples in changing environments [1]. One instantiation of the meta-learning is the task of Few-shot classification.
In this task the problem is formulated as assigning a class value to a query image, where classes are specified by one (or few) examples in a support (data) set (Figure 1). Here both the query image and the example images for each class in the support set have never been seen by the model during training. Moreover the content of the images in the query and support set may be new to the model.
Humans are very capable of inferring the aspect of the question in this setting. However, meta-learning algorithms are not specifically designed for this given the assumption of a single discriminating class description.
The goal of this project is to extend the capabilities of meta-learning models towards aspects-based meta-learning. One of the possible reasons for lack of a solution for this problem formulation may be the incompatibility of the traditional label-based supervision learning approaches for dealing with the variable aspects of high-dimensional data. However, much progress has been made in learning from other types of supervision such as psychometric testing [2]. Enabled with such techniques, the goal of the project is to develop a method that can address the challenge of Aspect-based meta-learning.
This project has been kicked off by a MSc project of Phuong Trinh. Her study and development can be used as a starting point for achieving the goal of aspect based few shot learning. The thesis will be available shortly.
Supervisor: Vlado Menkovski
[1] https://lilianweng.github.io/lil-log/2018/11/30/meta-learning.html
[2] Yin, L., Menkovski, V., & Pechenizkiy, M. (2020). Knowledge Elicitation using Deep Metric Learning and Psychometric Testing. arXiv preprint arXiv:2004.06353. arxiv link