Problem
With very few labeled examples, semi-supervised learners can produce unreliable pseudo-labels. Training on those labels can reinforce the initial errors.
Approach
DP-SSL learns labeling functions through multiple-choice learning instead of requiring hand-written expert rules. A label model reconciles their noisy, overlapping predictions into probabilistic labels for the unlabeled examples.
Evidence
Experiments cover four standard semi-supervised benchmarks. On CIFAR-10 with 40 labeled samples, the paper reports 93.82% annotation accuracy on unlabeled data and 93.46% test classification accuracy; these are different evaluation targets.
Connection
The work treats the reliability of supervision as a modeling problem. This is a methodological foundation for the site's interest in evidence quality, not a demonstrated recommendation-system result.
Sources
Bibliographic details
- Authors
- Yi Xu, Jiandong Ding, Lu Zhang, Shuigeng Zhou
- Venue
- NeurIPS 2021
- Status
- Published