Neural network model learning and application methods, devices, systems and storage media

摘要

PROBLEM TO BE SOLVED: To provide a method for accelerating neural network model learning in consideration of the influence of a past gradient on a gradient in a learning iteration. A method of learning a neural network model includes determining a constraint threshold range according to the calculation accuracy of the neural network model and the number of learning iterations, and constraining a weight gradient within the constraint threshold range. Therefore, if the inaccurate weight gradient is distorted due to quantization error, the gradient distortion is corrected by the gradient constraints, thereby ensuring that the trained network model achieves the expected performance. .. [Selection diagram] FIG. 11