Title:	A Generic Active Learning Framework for Deep Models
DNr:	Berzelius-2022-11
Project Type:	LiU Berzelius
Principal Investigator:	Linus Aronsson <linaro@chalmers.se>
Affiliation:	Chalmers tekniska högskola
Duration:	2022-01-25 – 2022-08-01
Classification:	10201
Keywords:

Abstract

The project is a part of my (WASP funded) PhD titled "A Generic Active Learning Framework for Deep Models". Active learning (AL) is a form of machine learning where the learning algorithm can interactively query for an annotation according to some objective in order to improve the learning in different ways. In general, different AL methods perform two operations for every annotation: i) selection which finds the next annotation according to some objective or acquisition function, and ii) update that applies the new annotation(s) and updates the model. Use cases include reducing data labeling efforts for supervised learning or finding more relevant clusters for a particular application for unsupervised learning. AL has been applied successfully to a number of tasks involving traditional (non-deep) machine learning models. With the development of advanced complex and deep models such as deep neural networks, both selection and update operations face several computational, methodological and theoretical challenges. In this project, we will develop general-purpose methods for AL of complex machine learning models. A recent approach to AL is to learn the acquisition function from data in order to achieve better performance and transferability between tasks. For example, the AL procedure could be viewed as a sequential decision process where reinforcement learning algorithms can be utilized to find good policies (i.e., acquisition functions). Most prior work on this does not use deep learning. The first project will therefore consist of investigating the idea of learning AL from data with the use of e.g. deep reinforcement learning applied to various deep learning tasks. The purpose of the project is to develop powerful AL strategies that can be used to intelligently select which samples to label in some large unlabeled dataset. The idea is to minimize the number of labels needed in order to achieve some performance of interest. Manually assigning labels to datasets can be very expensive in many cases which should highlight the importance of the project. However, note that the developed ideas may be quite general and could be applicable to other forms of AL applications. The software methods used will involve implementing various deep learning models in PyTorch and various deep reinforcement algorithms using libraries such as open-baselines3 (https://stable-baselines3.readthedocs.io/en/master/) which also uses PyTorch as their DL backend.