Compensation Method of Quantized Deep Learning Models for Edge Devices

Quantization is one of the optimization methods for developing deep learning models for edge devices. Through converting the floating-point into 8bit integer or even lower bitwidth, the model’s storage size can be reduced. As the rounding error exists during the quantization process, the model performance decreases. As a result, a method that can recover model performance is needed. In this research, a compensation method for improving the performance of quantized deep learning models is proposed, which make the quantized model can achieve equal or even better performance compared to the original floating-point model.