Work

Measuring and acting on uncertainty in deep neural networks: Selective prediction and calibrating confidence

Year: 2023
Type: dissertation
Abstract: Many machine learning (ML) models suffer from overconfidence and mismatch be- \ntween a model’s prediction and the likelihood of true correctness. Outside of the lab- \noratory setting, ML models usually operate within the context of larger systems. A \nmodel’s inability to detect when it is uncertain about its prediction can lead to critical \nissues. There are different approaches to measure and act on uncertainty in predictions \nwhen deploying ML models in the real-world. This thesis explores two in depth. The \nfirst approach is to integrate a reject option for ML models allowing them to abstain \nfrom making predictions when they are uncertain. We adopt the framework known as \nselective networks. However, optimizing selective networks is challenging due to the \nnon-differentiability of the binary selection function (the discrete decision of whether \nto predict or abstain). We propose an alternative framework to train selective networks \nthat employs the Gumbel-softmax reparameterization trick to enable selection within \nan end-to-end differentiable training framework. Our framework showed better perfor- \nmance on different classification and regression benchmarks and resulted in a reduced \ngap between the selection rates during training and test time. The second approach is \ntraining models to have well-calibrated confidences. Calibrated confidences are more \nrobust in the real-world as they are more representative of the true correctness like- \nlihood. We propose a framework that adapts knowledge distillation for training ML \nmodels to have better calibration in long-tailed recognition settings. Our approach \nachieves state-of-the-art results on various visual recognition benchmarks.
Language: English
Open Access status: green
OpenAlex ID: W7067273407
Cites: 80
Cited by:
Related to: 10