Improved Bayesian adaptive resonance classification method

Abstract

The invention discloses an improved Bayesian adaptive resonance classification method which comprises the following steps: constructing an RQ-BR-BAM model, introducing a rational quadratic kernel function and a Bayesian regularization overfitting suppression mechanism to preprocess data, and setting test parameters and RQ-BR-BAM model parameters; training an RQ-BR-BAM model through the data preprocessed in the step 2; and performing a classification experiment through the trained model and comparing the inference accuracy of the model. The generalization ability of the model is improved by introducing two types of over-fitting suppression mechanisms; a rational quadratic kernel function is introduced, the form and parameters of the function are changed, mapping from an input variable to afeature space is implicitly changed by adjusting the parameters of the function, the class cluster boundary is prevented from being highly sensitive to new data to prevent overfitting, the generalization ability of the model is improved, and logarithm is taken for a likelihood estimation item and a prior probability item. The bias of a prior probability item and a likelihood estimation item in theearly stage and the later stage of training is reduced to realize the stability convergence of the model.

Description

technical field [0001] The invention relates to the technical field of neural networks, and more specifically relates to an improved Bayesian adaptive resonance classification method. Background technique [0002] At present, the Bayesian Adaptive Resonance Theory MAP (BARTMAP), referred to as BAM, solves the problem of undefinable number of clusters and category diffusion, and achieves good classification performance and has been widely used. BAM itself still has unavoidable problems: 1) The calculation cost is too high: in the Bayesian clustering process, the Gaussian distribution function is defined to continuously calculate and update the covariance matrix to match and adjust the size and shape of the clusters, but the covariance matrix The calculation cost of the variance is too high; 2) The performance is unstable: the posterior probability that determines the classification effect depends on the likelihood estimation item and the prior probability item. When the dimen...

AI Technical Summary

Problems solved by technology

[0002] At present, the Bayesian Adaptive Resonance Theory MAP (BARTMAP), referred to as BAM, solves the problem of undefinable number of clusters and category diffusion, and achieves good classification performance and has been widely used. BAM itself still has unavoidable problems: 1) The calculation cost is too high: in the Bayesian clustering process, the Gaussian distribution function is defined to continuously calculate and update the covariance matrix to match and adjust the size and shape of the clusters, but the covariance matrix The calculation cost of the variance is too high; 2) The performance is unstable: the posterior probability that determines the classification effect depends on the likelihood estimation item and the prior probability item. When the dimension of the sample data is high, the likelihood estimation item becomes abnormal. Stable and difficult to achieve global convergence, which affects the classification performance of BAM

[0003] In response to the above problems, the researchers proposed the KernelBayesian Adaptive Resonance Theory Classification Model (KernelBayesian ARTMAP, KBAM), which introduced Kernel Bayesian rules and entropy-induced metrics to avoid a large number of covariance calculations and save computational costs. At the same time, the calculation process of the posterior probability item is avoided, and the stability of the model is improved, but the entropy-induced metric value weakens the potential of local information crossover to a certain extent, and the accuracy rate decreases during the small-sample training process. The generalization of the model Insufficient ability, resulting in a significant decline in classification performance

[0004] However, although BAM and KBAM can solve the problems of category diffusion caused by noise interference and instability of high-dimensional samples in the classification process, they still cannot handle small sample data in high dimensions well at this stage.

Images