Establishing Quantified Uncertainty in Neural Networks
link : https://github.com/mit-ll-responsible-ai/equine
Developed the initial protonet models for this effort.
Deep neural networks (DNNs) for supervised labeling problems are known to produce accurate results on a wide variety of learning tasks. However, when accuracy is the only objective, DNNs frequently make over-confident predictions, and they also always make a label prediction regardless of whether or not the test data belongs to any known labels.