On the application of quantization for mobile optimized convolutional neural networks as a predictor of realtime ageing biomarkers

摘要

In this paper we propose a mobile optimized deep learning network based on the VGG16 architecture. Compared to the classical approach, after training has been performed the model is converted to a quantized equivalent where 32 bit floating point operations are exchanged for 8 bit ones. This reduces the strain on mobile memory and local caches while simultaneously reducing the computational complexity and energy requirement of the entire deep learning model. Aggregated testing has been performed to validate the complexity hypothesis and the quantized model has been compared to the original model in terms of accuracy. The results show that for a modest decrease in accuracy, the quantized model takes up 75% less disk space and through the 8 bit operations the computational complexity is reduced, showing a load and inference speed up of 3 - 4 times faster than the original model.