Learning to learn and extend with meta-learning

Traditionally, machine learning algorithms are designed to receive a fixed dataset as an input and after training the model will stay the same. However, in real world applications, the machine learning models often need to evolve with a constantly changing dataset that increases with new training examples. The main challenge is to tackle catastrophic forgetting; that is, avoid the classifier to only optimise towards new classes. There are different ways to do so, including controlling gradient updates by considering both old and new classes prediction, using in-memory samples to replay, and employing multi-classifiers for different tasks. In this project, we look into a new way, using meta-learning to learn how the classifier evolve, and is impacted by the learning of new classes. The research question is: can we use that information to improve learning quality and reduce catastrophic forgetting? One way to do is to use meta-learner to capture and track such information and learn update policy on the classifier.

Supervisors

Artefact(s)

  1. Apply the meta-learning approach [1] to sensor data based human activity recognition. The datasets will be given. 

 

 

Background

[1] Jiawei Wu, Wenhan Xiong, William Yang Wang. Learning to Learn and Predict: A Meta-Learning Approach for Multi-Label Classification. 2019. https://arxiv.org/abs/1909.04176