Incorporating Task-Agnostic Information in Task-Based Active Learning Using a Variational Autoencoder

When building datasets for machine learning, it is often much easier and less expensive to collect datathan to label it, resulting in large pools of unlabeled data. Active learning (AL) is a subfield of machine learning that focuses on choosing which unlabeled points to label next for a given dataset and task, with the core assumption that labeling certain points will result in higher performance models than other points. Standard AL approaches identify informative samples by querying a trained task model. Task-agnostic AL approaches ignore the task model and instead makes selections based on separately defined properties of the dataset. We seek to combine these approaches and measure the contribution of incorporating task- agnostic information into task-focused AL. We use a ResNet classifier as our task model and experiment across two AL utility functions with and without added information from a variational autoencoder (VAE).