Multi-output regression network and learning method

Abstract

The invention provides a multi-output regression network for learning an input-output relation, a method for learning the input-output relation based on the multi-output regression network, and a computer readable storage medium. The multi-output regression network comprises a plurality of input nodes, a plurality of intermediate nodes and a plurality of output nodes, wherein the input nodes are used for receiving input data (xi); a non-linear network layer is formed between the intermediate nodes and the input nodes; the non-linear network layer maps the input data (xi) to the intermediate nodes through non-linear mapping; the intermediate nodes apply a predetermined function to the input data (xi) subjected to the non-linear mapping; a linear network layer is formed between the output nodes and the intermediate nodes; and the linear network layer maps the input data applied with the predetermined function to the output nodes through linear mapping and outputs the data as output data(yi). The multi-output regression network learns the relation between the input data (xi) and the output data (yi) by performing joint optimization on mapping parameters (W) of the non-linear mappingand mapping parameters (S) of the linear mapping under the condition of giving the input data (xi) and the output data (yi).

Description

technical field [0001] The invention relates to the field of machine learning, in particular to a multi-output regression network for learning input-output relationship and a method for learning input-output relationship. Background technique [0002] With the explosion of deep learning, neural networks have once again become a hot spot of attention. In the field of machine learning and data mining, kernel approximation methods and multi-output regression models are currently used to describe the relationship between input and output. [0003] The kernel approximation method is to find a mapping function based on Bechner's theorem, which just corresponds to a translation-invariant kernel function. The basic principle of the kernel approximation method is that a positive definite, translation-invariant kernel function is a distributed Fourier transform. Through the decomposition of the basis function of the Fourier transform, the mapping function can be approximately express...

AI Technical Summary

Problems solved by technology

However, in most practical applications, the assumption that the regression coefficients exist in a low-dimensional manifold space in the correlation assumption model based on prior knowledge is not valid, and in different applications, it cannot be guaranteed that a linear subspace can be shared. These two points greatly limit the application of this method in real data, and this special assumption will seriously affect the system performance

Also, linear regression models do not handle nonlinear relationships between input and output well

Images