This paper addresses the quantization of weight coefficients in trained neural networks to achieve more compact representations. To mitigate the degradation of the input-output relationship in neural networks due to quantization, we introduce the noise-shaping quantization method. This method quantizes coefficients and distributes the resulting quantization error over coefficients not yet quantized. In this paper, we proposed an approach to adjust the magnitude of the quantization error using training data. The effectiveness of the proposed method was then validated through an image classification problem. Finally, we related the quantization problem of neural network weight coefficients to the quantization problem of time signals in dynamic systems and discussed its distinctions from previous quantization techniques.

Mass-produced control systems allow tolerances for each component. Therefore, the model parameters of the controlled object also have a certain tolerance, and the controller design is generally designed based on the nominal value considering the perturbation. As a result, the target value response characteristics also have a certain range of variation. The structure uses an ILQ controller and a precompensator, and the experimental data obtained in the inspection process etc. A design method that adjusts the controller using the FRIT method and restores it to the nominal target value response characteristic (this paper refers to this as trimming) is proposed. We propose this method and demonstrate its effectiveness using a numerical example.