Developing novel strategies for integrating Active and Deep Learning for Artificial Intelligence
Data play a critical role in the further development of Artificial Intelligence. More precisely annotated data are required to train today’s Artificial Intelligence algorithms. Data annotation, however, is a time-consuming and expensive task that is to a certain degree subjective. Thus, novel approaches to accelerate the annotation process with fewer user interactions are needed.
A promising solution is Active learning: a learning methodology that addresses the problem of learning from not-annotated data efficiently with minimum input from the user. It reduces the number of labelled data needed to train a model by leveraging correctly classified data provided by the user in a feedback loop. To minimise user interaction in this learning process, information is only requested for those samples, which contribute most to the learning process. Active Learning has been successfully applied in the fields of image categorisation, natural language processing, image classification, sequence labelling, and image segmentation.
Combining Active Learning and Deep Learning
Deep Learning is another machine learning method that has progressed machine learning significantly. However, it relies heavily on labelled input data. Whenever unlabelled data is used, user input in the form of active learning is needed. This is especially true as machine learning is applied to more and more types of data without readily available annotation. Combining Active and Deep Learning is challenging, since Deep Learning methods are rather slow and require many labelled input samples. Active Learning, however, requires fast learning methods that need only a few labelled data points.
This project investigates this challenge in depth and develops novel strategies for integrating Active and Deep Learning in a joint “Deep Active Learning” framework. The focus is put on object detection and classification for unlabelled data. The successful combination of both methods could progress machine learning from unlabelled data substantially.