提升分类受限玻尔兹曼机性能的策略

Strategy to improve performance of classification restricted Boltzmann machine

为提高分类受限玻尔兹曼机(classification restricted Boltzmann machine,ClassRBM)有限的学习能力,提出一种基于重构误差的学习助推策略,提升ClassRBM的分类性能。重构误差是模型生成的数据与原始数据之间的差异,其会影响模型的性能。通过设置不同的重构误差阈值,选择重构误差超过阈值的原始数据对强化模型进行训练。测试时,统计测试数据集中被ClassRBM分错,且重构误差超过阈值的测试数据,如果存在这样的测试数据,错分数据采用强化模型的分类结果。在不同数据集上的测试结果表明,提出策略能提升ClassRBM的性能。

To increase the limited learning ability of classification restricted Boltzmann machine(ClassRBM),a learning boosting strategy based on reconstruction error was proposed to improve the classification performance of ClassRBM.The reconstruction error was the difference between the data generated by the model and the original data.This difference affected the performance of the model.Different reconstruction error thresholds were set up,and the enhanced model was trained by the original data whose reconstruction error exceeded the threshold.In testing,the test data which was misclassified using ClassRBM and its reconstruction error exceeded the threshold were counted.If such data existed,the final result of the data was the classification result of the enhanced model.The results on different data sets show that the proposed strategy can improve the performance of ClassRBM.