Document Type

Conference Proceeding

Publication Date

9-2023

Publication Title

CARLA 2023: The Latin America High Performance Computing Conference

Pages

1-11

Publisher Name

Springer

Publisher Location

Switzerland

Abstract

Deep learning (DL) has become a cornerstone for advancements in computer vision, yielding models capable of remarkable performance on complex visual tasks. Despite these achievements, there remains a critical need for accurate uncertainty estimations, especially when models encounter out-of-distribution (OOD) inputs. Addressing this, our research focuses on the implementation and evaluation of uncertainty techniques in two prominent DL architectures: Convolutional Neural Networks (CNN) and Vision Transformers (ViT). These architectures were applied specifically to computer vision tasks, utilizing the MNIST and ImageNet-1K datasets for our evaluations.

High-Performance Computing (HPC) platforms, pivotal to this research, were employed to assess these techniques. The traditional Polaris supercomputer, equipped with AMD EPYC processors and NVIDIA A100 GPUs, is evaluated alongside cutting-edge AI accelerators: Cerebras CS-2 and SambaNova DataScale. These platforms, with their unique architectures, presented varying efficiencies in the application of uncertainty estimation. Our findings elucidate the computational advantages and challenges associated with each, providing invaluable insight for researchers.

Furthermore, this paper offers practical considerations when employing HPC systems for uncertainty estimations in DL. From computational setup to architectural nuances, the insights garnered pave the way for future research aiming to integrate algorithms and hardware for robust model predictions in computer vision and other areas of DL.

Comments

Author Posting © Springer, 2023. Accepted paper at CARLA 2023. This work was funded by Department of Energy.

Available for download on Sunday, December 01, 2024

Share

COinS