End-to-End Weak Supervision

• DOI: 10.48550/arxiv.2107.02233
• Publication Date: 2021-01-01
• Authors: Salva Rühling Cachay; Benedikt Boecking; Artur Dubrawski
• Keywords: Robustness; End-to-end principle; Training set; Ground truth; Labeled data

Aggregating multiple sources of weak supervision (WS) can ease the data-labeling bottleneck prevalent in many machine learning applications, by replacing tedious manual collection ground truth labels. Current state art approaches that do not use any labeled training data, however, require two separate modeling steps: Learning a probabilistic latent variable model based on WS -- making assumptions rarely hold practice followed downstream training. Importantly, first step does consider performance model. To address these caveats we propose an end-to-end approach for directly maximizing its agreement with labels generated reparameterizing previous posteriors neural network. Our results show improved over prior work terms end test sets, as well robustness to dependencies among sources.