Quantization Method Based on Weight Compression and Format Simplification

Abstract

With the combination of artificial intelligence and the Internet of Things, related technology applications become more diversified. Artificial intelligence is no longer only used in cloud servers as in the past but is available in specific fields. Therefore, many artificial intelligence applications are currently imported into the embedded system architecture. The embedded system has storage space, energy consumption, and computing performance limitations. If the training model is directly embedded in the system, the embedded system does not operate normally. Therefore, we propose a novel quantization algorithm to simplify the data format in the model. At the same time, the location with a lower density is found by using the normal distribution detection in statistics to determine the weight distribution, and the value of the adjacent location replaces that of the location. The results show that even if the data format is modified, the feature does not disappear. In the case of using fewer testing resources, there are prominent features that increase the identification accuracy.