8 Bit Quantized Neural Network

– Efficient deployment of deep neural networks on edge devices is constrained by limited computational resources, memory, and energy availability. While full-precision (FP32) models provide high accuracy, their resource demands hinder practical edge implementation. This paper evaluates 8-bit integer (INT8) quantization using Quantization-Aware Training (QAT) on the ResNet-18 architecture with the CIFAR-10 dataset. By modeling quantization effects during training, the approach reduces the accuracy loss typically associated with lower precision. Experimental results show that the INT8 model achieves 94.77% accuracy, closely matching the FP32 baseline of 94.79%, while significantly improving efficiency in terms of model size and computation. These findings demonstrate that INT8 quantization offers an effective trade-off between performance and resource utilization for real-time edge AI applications.