Fidelity-Weighted Learning (arXiv)

Deghani et al. propose a new approach for semi-supervised learning with weak supervision. Their framework trains a student network on weakly annotated data. They then train a Gaussian Process-based teacher on gold data using the student's representations and use the teacher to estimate the confidence of the weakly supervised examples. Finally, the student is fine-tuned by incorporating on the weakly supervised data by incorporating the confidence estimates.


Want to receive more content like this in your inbox?