Neural network weight discretizing method, system, device, and readable storage medium

Abstract

There are provided a neural network weight discretizing method, system and device, and a computer readable storage medium. The method includes acquiring a weight value range and a number of discrete weight states, the weight value range referring to a range of discrete weight values consisting of a maximum weight value of a current time step and a minimum weight value of the current time step, and the number of discrete weight states referring to the quantity of discrete weight states. The method also includes acquiring a weight state of a previous time step and a weight increment of the current time step and acquiring a state transfer direction by using a directional function according to the weight increment of the current time step. The method also includes acquiring a weight state of the current time step according to the weight state of the previous time step, the weight increment of the current time step, the state transfer direction, the weight value range and the number of discrete weight states. The method ensures that the weight value is always constrained in the same discrete-state space without the need for storing an additional virtual continuous-state implicit weight. On the premise that the computation performance of the neural network is ensured, the consumption of storage space is greatly reduced, and the computation complexity is reduced.

Dependencies are inferred from the claim text.