Incorporating Task-Agnostic Information in Task-Based Active Learning Using a Variational Autoencoder
Curtis Godwin
Meekail Zain
Nathan Safir
Bella Humphrey
Shannon P Quinn
It is often much easier and less expensive to collect data than to
label it. Active learning (AL) (settles2009active) responds to this issue by selecting
which unlabeled data are best to label next. Standard approaches
utilize task-aware AL, which identifies informative samples based on
a trained supervised model. Task-agnostic AL
ignores the task model and instead makes selections based on learned
properties of the dataset. We seek to combine these approaches and
measure the contribution of incorporating task-agnostic information
into standard AL, with the suspicion that the extra information in
the task-agnostic features may improve the selection process.
We test this on various AL methods using a ResNet
classifier with and without added unsupervised information from a
variational autoencoder (VAE). Although the results do not show a
significant improvement, we investigate the effects on the
acquisition function and suggest potential approaches for
extending the work.
active learning, variational autoencoder, deep learning, pytorch, semi-supervised learning, unsupervised learning
DOI10.25080/majora-212e5952-011