Promises and Pitfalls of Threshold-based Auto-labeling
This work addresses the problem of dataset quality assurance for practitioners using auto-labeling, though it is incremental as it provides the first formal analysis of an existing method.
The paper tackles the challenge of ensuring quality in threshold-based auto-labeling systems by deriving sample complexity bounds for the required human-labeled validation data, showing that seemingly bad models can accurately label large unlabeled data chunks but may require prohibitively large validation sets.
Creating large-scale high-quality labeled datasets is a major bottleneck in supervised machine learning workflows. Threshold-based auto-labeling (TBAL), where validation data obtained from humans is used to find a confidence threshold above which the data is machine-labeled, reduces reliance on manual annotation. TBAL is emerging as a widely-used solution in practice. Given the long shelf-life and diverse usage of the resulting datasets, understanding when the data obtained by such auto-labeling systems can be relied on is crucial. This is the first work to analyze TBAL systems and derive sample complexity bounds on the amount of human-labeled validation data required for guaranteeing the quality of machine-labeled data. Our results provide two crucial insights. First, reasonable chunks of unlabeled data can be automatically and accurately labeled by seemingly bad models. Second, a hidden downside of TBAL systems is potentially prohibitive validation data usage. Together, these insights describe the promise and pitfalls of using such systems. We validate our theoretical guarantees with extensive experiments on synthetic and real datasets.