Skip to main content

Deep Active Learning with Rationales

Primary supervisor

Lan Du


  • Dr Ming Liu

The performance of deep neural models rely on large amounts of labeled data, however, most data remain unlabeled in the real world scenario. While annotating data is expensive and time consuming, active learning seeks to choose the most appropriate and worthwhile data for human annotation. It is noticed that humans give labels to some specific data with some labeling reasons or rationales,  which are often existing in the data.  The goal of this research is to develop effective deep active learning techniques with rationales.

Student cohort

Double Semester


This project will investigate various query strategies in deep active learning and design novel methods to incorporate the rationales into the active learning cycle. So that the querying process is equipped with a certain level of explainability. The expected outcome is new deep active learning algorithms and effective mechanisms to leverage human rationales.

Required knowledge

  • Proficiency in Python programming, with experience in Pytorch/Tensorflow
  • Basic knowledge of machine learning
  • Knowing active learning is a plus but not necessary.