On the Effect of Quantization on Deep Neural Networks Performance

Neural network Quantization has emerged as an important technique for reducing the computational cost of deep neural networks (DNNs) and deploying them on resource-constrained devices. However, DNN models can exhibit vulnerabilities when exposed to various types of noise in real-world applications. While evaluating the impact of quantization on model performance is crucial, research on this topic is notably limited, often neglecting established principles of performance evaluation and consequently yielding inconclusive findings. To bridge this gap, we evaluated quantized models’ performance under various noise types, including data perturbation, model parameter perturbation, and adversarial attacks, using the Tiny ImageNet dataset. The empirical results from this comprehensive evaluation present a valuable understanding of how quantized models perform across diverse scenarios, particularly when compared to the performance of the original models.