基于稀疏降噪自编码器的深度置信网络

发布时间：2018-06-18 12:34

本文选题：深度置信网络 + 受限玻尔兹曼机　；参考：《计算机应用》2017年09期

【摘要】：传统的深度置信网络(DBN)采用随机初始化受限玻尔兹曼机(RBM)的权值和偏置的方法初始化网络。虽然这在一定程度上克服了由BP算法带来的易陷入局部最优和训练时间长的问题,但随机初始化仍然会导致网络重构和原始输入的较大差别,这使得网络无论在准确率还是学习效率上都无法得到进一步提升。针对以上问题,提出一种基于稀疏降噪自编码器(SDAE)的深度网络模型,其核心是稀疏降噪自编码器对数据的特征提取。首先,训练稀疏降噪自编码;然后,用训练后得到的权值和偏置来初始化深度置信网络;最后,训练深度置信网络。在Poker Hand纸牌游戏数据集和MNIST、USPS手写数据集上测试模型性能,在Poker Hand数据集下,方法的误差率比传统的深度置信网络降低46.4%,准确率和召回率依次提升15.56%和14.12%。实验结果表明,所提方法能有效地改善模型性能。
[Abstract]:The traditional depth confidence network (DBN) initializes the weights and biases of constrained Boltzmann machine (RBM) at random. Although this can overcome the problems of local optimization and long training time brought by BP algorithm, random initialization still leads to the difference between network reconstruction and original input. This makes the network not only in accuracy but also in learning efficiency can not be further improved. In order to solve the above problems, a kind of depth network model based on sparse noise reduction self-encoder (SDAE) is proposed, the core of which is the feature extraction of data by sparse de-noising self-encoder. First, the sparse denoising self-coding is trained; then, the depth confidence network is initialized with the weight and bias obtained by the training; finally, the depth confidence network is trained. The performance of the model is tested on the Poker hand and MNIST / USPS handwritten data sets. Under the Poker hand dataset, the error rate of the method is 46.4% lower than that of the traditional depth confidence network, and the accuracy and recall rate are increased by 15.56% and 14.12% respectively. Experimental results show that the proposed method can effectively improve the performance of the model.
【作者单位】：广东工业大学计算机学院;广东建设职业技术学院现代教育技术中心;西蒙弗雷泽大学影像技术实验室;
【基金】：国家自然科学基金资助项目(61300107) 广东省自然科学基金资助项目(S2012010010212) 广州市科技计划资助项目(201504301341059,201505031501397)~~
【分类号】：TP18

【相似文献】