Calculation device and method applicable to neural network activation function

Abstract

The invention relates to a calculation device and method applicable to a neural network activation function. The calculation device applicable to the neural network activation function comprises a searching unit and a calculation unit, wherein the searching unit is used for determining a corresponding linear function parameter according to an activated function and an activated function variable input value and the calculation unit is used for using the input value and a corresponding linear function parameter to perform calculation. The linear function parameter can perform optimization according to concavity and convexity of the activated function.

Description

technical field [0001] The invention relates to the computing field, in particular to a computing device and method suitable for neural network activation functions. Background technique [0002] With the development of science and technology, neural network, as a machine learning technology that simulates the human brain, has always been a hot spot in the field of artificial intelligence research, and has been widely used in many fields, such as image recognition. One of the key points of using neural networks to solve practical problems is to use activation functions to provide nonlinear modeling capabilities for neural networks. Usually, activation functions appear in each layer of neural networks, especially in certain algorithms. Higher, such as the sigmoid function. Therefore, when performing calculations on the activation function, its calculation speed and calculation energy consumption directly restrict the calculation efficiency of the entire neural network. [0...

AI Technical Summary

Problems solved by technology

[0003] Traditional neural network computing devices generally refer to general-purpose arithmetic logic units. When performing calculations on activation functions, the resource utilization rate is low and the execution speed is slow.

Especially with the development of various portable electronic devices, the circuit hardware overhead of the above-mentioned devices is too large to meet the needs of application scenarios with small size and low power consumption.

Images