Active Learning and Pseudo-Labeling for Machine Learning

Summary

The document explains semi-supervised and active learning, then describes how a classifier can use unlabeled examples to guide which labels to request. It outlines assumptions that support learning from unlabeled data, including nearby points tending to share labels, clusters tending to contain similar labels, and data lying on a lower-dimensional manifold. It also describes pseudo-labeling, in which a model or proximity rule assigns provisional labels that are combined with the known labels for training.

For active learning, the article presents membership-query, stream-based, and pool-based sampling, with uncertainty, margin, and entropy as ways to rank candidate examples. It discusses combining query measures and illustrates the workflow with classification models and a library built around scikit-learn. The reported experiments compare active-learning approaches with passive training, including committees of learners; the author finds no clear, reliable improvement and notes that results vary. The tests are not evidence of a trading edge, and the conclusion stresses the need for careful feature and label preparation. Performance depends on the dataset, query strategy, and evaluation setup.

Key ideas

• Semi-supervised learning uses unlabeled examples under assumptions about the data’s structure and label distribution.
• Pseudo-labeling assigns provisional labels to unlabeled examples and adds them to the training data.
• Active learning selects examples for labeling according to an informativeness measure.
• Uncertainty, margin, and entropy sampling are common ways to rank candidate examples.
• The reported experiments show no consistent advantage over passive learning, and the author highlights the cost of added training work.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.