Prior Learning Based On Kernelized Stein Discrepancy And Its Application

Given data set,estimating the true probability density distribution of data is a fundamental problem in machine learning and statistics.Traditional Bayesianbased methods,such as maximum likelihood estimation and maximum a posteriori estimation.For maximum likelihood-based The model comparison and parameter estimation problem with the maximum posterior is only applicable to the standardized probability density function.The true probability density distribution of the data is not standardized,and there are often difficult to solve partition functions in the prior distribution,so these cannot be used directly.Method for model comparison or parameter estimation.In this paper,the non-standardized estimation method Stein difference is used to describe the matching degree between the data and the non-standardized model.The problems existing in other methods are avoided.There is noise in signals in practical applications,and traditional noise reduction methods cannot well describe the prior distribution of complex signals.In this paper,the method based on Product of Expert is used,and a single expert uses the Student-t distribution,which can describe the prior distribution of complex signals.The probability density function of the heavy-tailed distribution of continuous random variables.It can describe the sparsity of the signal.Firstly,the signal is divided into overlapping frames,and its frequency domain representation is obtained by short-time Fourier transform,and the signal is denoised in the frequency domain to obtain the denoised signal representation.That is,the prior distribution of the signal with parameters.Aiming at the problem of parameter estimation.In this paper,the KSD minimization is used to learn the PoE model parameters,and it is used for the signal noise reduction task,thus the KSD-PoE noise reduction method is proposed,and it is applied to the heart sound signal The experimental results show the effectiveness of this method.