Trustworthy deep learning acceleration with customizable design flow automation
File(s)
Published version
Author(s)
Type
Conference Paper
Abstract
In recent years, deep learning has brought the development of accurate and complex models across various domains. Deploying these models efficiently on resource-constrained platforms while maintaining high accuracy and trustworthiness, however, remains a critical challenge. This paper introduces an automated framework for optimizing trustworthy deep learning by enabling trade-off between three metrics: computational efficiency, trustworthiness, and predictive accuracy. Traditional compression techniques such as pruning and scaling reduce computational complexity but can compromise model calibration and uncertainty quantification, which is critical for safety-critical applications. To address this challenge, we integrate Monte Carlo Dropout (MCD) for Bayesian Convolutional Neural Networks (BayesCNNs) and propose an automated Design Space Exploration (DSE) approach driven by Bayesian Optimization to identify Pareto-optimal configurations. Our framework dynamically tunes pruning rates, dropout probabilities, and other parameters to achieve Pareto-optimal trade-offs between accuracy, efficiency, and uncertainty estimation. Two BayesCNN architectures are evaluated to demonstrate that our approach can systematically optimize deep learning models for trustworthiness and efficiency. Our results show that no single configuration is optimal across all metrics, demonstrating the need to automate and customize co-optimization strategies. Compared to state-of-the-art FPGA implementations, our optimized design achieves up to 7.67× faster inference and 12.8× higher energy efficiency while maintaining well-calibrated uncertainty estimates.
Editor(s)
Ueno, T
Abdelhadi, A
Koch, D
Osana, Y
Sato, Y
Date Issued
2025-05-25
Date Acceptance
2025-05-01
Citation
HEART '25: Proceedings of the 15th International Symposium on Highly Efficient Accelerators and Reconfigurable Technologies, 2025, pp.1-13
ISBN
9798400714320
Publisher
Assoc Computing Machinery
Start Page
1
End Page
13
Journal / Book Title
HEART '25: Proceedings of the 15th International Symposium on Highly Efficient Accelerators and Reconfigurable Technologies
Copyright Statement
© 2025 Copyright held by the owner/author(s). This work is licensed under a Creative Commons Attribution 4.0 International License (https://creativecommons.org/licenses/by/4.0/).
License URL
Identifier
10.1145/3728179.3728198
Source
15th International Symposium on Highly Efficient Accelerators and Reconfigurable Technologies-HEART
Subjects
Computer Science
Computer Science, Hardware & Architecture
Computer Science, Theory & Methods
Science & Technology
Technology
Publication Status
Published
Start Date
2025-05-26
Finish Date
2025-05-28
Coverage Spatial
Kumamoto, Japan
Date Publish Online
2025-05-25