Cambrian neural network model quantization training method patent authorized

The Qichacha APP shows that recently, the patent for "a quantitative training method, device, and equipment for a neural network model" of Hanzhi Information Technology Co., Ltd., a wholly-owned subsidiary of Cambricon, has been authorized. The patent abstract of Qichacha shows that the method includes: obtaining the first input data and parameters of the layer to be quantified in the forward propagation process; quantifying the first input data and parameters separately to obtain quantified first input data and quantified first parameters; performing operations on the quantified first input data and quantified first parameters to obtain the first output data; determining whether the previous layer of the layer to be quantified depends on the first input data during backpropagation; if yes, saving the first input data. In the implementation of this application example, by selecting different quantization methods for different scenarios, the computational workload during the training process of the neural network model can be reduced, and memory space can be saved.