Neural network quantification method for solving regression problem

The invention relates to the technical field of neural networks, and discloses a neural network quantification method for solving a regression problem. The method comprises the steps of firstly obtaining a training data set; then designing a nonlinear activation function which is easy to realize by...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: MO PINGHUI, TAN ZILING, LIU JIE, ZHAO ZHUOYING
Format: Patent
Sprache:chi ; eng
Schlagworte:
Online-Zugang:Volltext bestellen
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:The invention relates to the technical field of neural networks, and discloses a neural network quantification method for solving a regression problem. The method comprises the steps of firstly obtaining a training data set; then designing a nonlinear activation function which is easy to realize by hardware; pre-training a 32-bit floating point type full connection layer neural network model based on the activation function; calling a pre-training model, quantizing the floating point type weight value into a form of integer power of 2, and performing fixed-point quantization on the floating point type bias value, floating point type input and output of each layer and input and output of an activation function; and finally, training the quantization network model by using back propagation and gradient descent algorithms. The shift summation operation is adopted to replace multiplication in the neural network, and a circuit is adopted to realize a simple activation function, so that the fitting precision of the