Abstract:
|
Classifiers have found widespread use for automating labor-intensive data collection in scientific pipelines. However, many scientific analyses involve downstream statistical inference that relies on the results of classifier predictions from the initial data collection. For such statistical inference to be valid, classifier predictions must generalize equally well across different experimental conditions. In this talk, we consider label shift as one assumption that allows for generalizable predictions, and we propose bootstrap methods for valid inference under label shift. We illustrate our methods with an application to live cell imaging data.
|