Evaluation of Novel AI Architectures for Uncertainty Estimation

Deep learning (DL) has advanced computer vision, delivering impressive performance on intricate visual tasks. Yet, the need for accurate uncertainty estimations, particularly for out-of-distribution (OOD) inputs, persists. Our research evaluates uncertainty in Convolutional Neural Networks (CNN) and Vision Transformers (ViT) using the MNIST and ImageNet-1K datasets. Using High-Performance (HPC) platforms, including the traditional Polaris supercomputer and AI accelerators like Cerebras CS-2 and SambaNova DataScale, we assessed the computational merits and bottlenecks of each platform. This paper delineates key considerations for using HPC in uncertainty estimations in DL, offering insights that guide the integration of algorithms and hardware for robust DL applications, especially in computer vision.