Hardware-In-The-Loop Progressive Compression and Dynamic Early-Exit Cnn for Edge Devices

View at publisher Summarise with VIRAT
Authors: Rajamohan R, Rahul J, Sinthuja P M
Cited by: 0
References: 17

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Deploying deep neural networks on resource-constrained edge devices requires hardware-aware adaptive compression. We introduce a multi-objective hardware-in-the-loop framework that optimizes accuracy, latency, energy, and memory jointly through: (i) hardware-specific cost-model-guided adaptive multi-granularity pruning, (ii) hierarchical knowledge distillation with latency-energy regularization, and (iii) energy-cognizant dynamic early-exit mechanisms with reinforcement learning-driven optimization. Our framework is verified on the Raspberry Pi 4 using ResNet-18 and MobileNetV2 on CIFAR-10/100, achieving up to 48.3 % latency reduction, 42.7 % memory savings, and 42.0% energy reduction at only 0.9 % accuracy degradation. Our reinforcement learning-driven optimized exit policy achieves$1.93 \times$speedup at an accuracy of 92.2 %.:

Fields of Study & Subjects

Advanced Neural Network Applications IoT and Edge/Fog Computing Parallel Computing and Optimization Techniques

Publication details

Journal / source
International Conference on Communication, Computing and Electronics Systems (ICCCES)
Pages
588–594
Publication date
21 January 2026
Language
EN