Workshops Archive

Optimized Uncertainty Estimation for Vision Transformers: Enhancing Adversarial Robustness and Performance Using Selective Classification

Workshop: 1st Workshop on Enabling Predictive Science with Optimization and Uncertainty Quantification in HPC

Authors: Erik Pautsch (Loyola University, Chicago); John LI (University of California, San Diego; Argonne National Laboratory (ANL)); Silvio Rizzi (Argonne National Laboratory); George Thiruvathukal (Loyola University, Chicago); and Maria Pantoja (California Polytechnic State University, San Luis Obispo)

Abstract: Deep Learning models frequently produce high-confidence softmax outputs for out-of-distribution (OOD) inputs, which would ideally be classified as "I don't know". To enhance our model's trustworthiness, we incorporate selective classification, which entails abstaining from predictions in situations of doubt. This approach requires initial uncertainty estimation. Subsequently, instead of offering a singular prediction, we provide a distribution over predictions, enabling users to discern if the model is trustworthy or if consultation with a human expert is necessary. In this paper, we assess uncertainty in two baseline models: a Convolutional Neural Network (CNN) and a Vision Transformer (ViT). Leveraging these uncertainty values, we minimize errors by refraining from predictions during high uncertainty. Additionally, we evaluate these models across various distributed architectures.

Back to 1st Workshop on Enabling Predictive Science with Optimization and Uncertainty Quantification in HPC Archive Listing

Back to Full Workshop Archive Listing