PQDE: Comprehensive Progressive Quantization with Discretization Error for Ultra-Low Bitrate MobileNet towards Low-Resolution Imagery

In deep learning, quantization is employed to tackle deployment challenges of neural networks in resource-limited environments like mobile and edge devices. Traditional full-precision (32-bit floating-point) models, while effective, are restricted by their high memory and computational demands, limiting their use in devices with constrained computational power and resources. To address this problem, we present a neural network quantization methodology that is primarily geared to-wards resource-constrained devices and inputs. Our methodology focuses on optimizing network performance for resource-limited settings, featuring a unique forward quantization function. This function employs the Minimize Discretization Error (MDE) technique to reduce information loss during quantization, particularly targeting near-zero weights, while maintaining computational efficiency and model accuracy. Additionally, we integrate the Arctangent Soft Round (ASR) method in the forward process to further smooth the data in low-bit quantization scenarios. Finally, we design a progressive quantization method, progressively transitioning from full precision to low bits, stabilizing the network at each quantization level. Tested on a resource-efficient variant of MobileNetV2 and low-resolution input data (CIFAR10/100), our method surpasses most contemporary techniques in terms of lightweight model performance. Through progressive quantization, our 4-bit quantized model even exceeds the accuracy of its full-precision counterpart as evidenced by our ablation studies.