.png){kind=icon}
.png){kind=icon}
Return
Companding quantization for deep neural networks
.png){kind=icon}
PRE
AI
DOI：10.1016/j.asoc.2025.113979.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
• This work introduces μ-law companding for post-training quantization, achieving non-uniform, asymmetric quantization optimized with Optuna. • VGG19 and AlexNet with CIFAR-10/100 datasets are evaluated demonstrating minimal accuracy loss ( <1 %) for 4-bit quantized models. • The “all mixed μ" configuration yielded the best performance, even surpassing full-precision model accuracy in some cases. • This work employs input normalization and dynamic μ adjustments to enhance quantization robustness and efficiency. • We show significant improvements over traditional methods, enabling efficient deployment on resource-constrained devices.