基于条件对数似然的BP神经网络多类分类器

Multi-class BP Neural Network Classifier Based on the Conditional Log-Likelihood

BP神经网络分类器存在收敛速度慢的缺陷,为了提高分类器性能,针对这一缺陷对BP算法进行改进.提出将条件对数似然(CLL)准则融入到监督性BP神经网络多类型分类过程中,利用CLL的可分解性优势,计算测试样本的条件概率,在误差反向传播时利用条件概率对权值进行相应的加权降权操作,简化误差反馈过程中的计算量.在实验中对改进算法的收敛速度和准确率进行了测试,说明了该算法的有效性及实用性.

BP neural network classifier has a slowly convergence rate, in order to improve the performance of the classifier, there is an improvement in BP algorithm for the problem. The Conditional Log-Likelihood （CLL） is applied into the supervisory neural network classification for the multi-class selection. By using the decomposability of CLL, calculate the conditional probability of the test samples. In the error back-propagation process, increasing or reducing the corresponding weights by using the conditional probabilities, which can simplify the computation in the process of error feedback. In the paper, we test the convergence speed and accuracy for the improved algorithm in the experiment. It illustrates the effectiveness and the practicality of the algorithm.