Journal of Engineering and Technology Research
· Volume Volume 1 · Issue Issue 01
Quantized Neural Network Architectures for Low-Latency Distributed Inference on Heterogeneous Edge Systems
Deploying modern deep neural networks across resource-constrained edge computing environments requires balancing model inference latency with accuracy and memory bandwidth. In this study, we propose a structured hybrid pruning and mixed-precision integer quantization framework tailored for heterogeneous edge clusters. Comprehensive empirical evaluations demonstrate a 3.4× reduction in inference latency and a 68% decrease in memory footprints across multi-node testbeds, while preserving over 99.2% of full-precision classification accuracy across benchmark datasets.