PPLS与稀疏鉴别流形正则化的双模型协同宽度神经网络

Dual Model Collaborative Broad Neural Network Based on PPLS and Sparse Discriminative Manifold Regularization

宽度神经网络(broad neural networks,BNN)被认为是继深度神经网络之后的一种主流机器学习算法,然而BNN没有考虑数据不确定性及局部几何结构信息。为此,提出概率偏最小二乘(probabilistic partial least square,PPLS)与稀疏鉴别流形正则化的双模型协同宽度神经网络建模方法。该方法首先使用PPLS对BNN输入特征以及增强特征构成的高维数据提取低维隐藏变量,消除数据不确定信息以及冗余特征;基于稀疏表示方法自适应构建样本局部与非局部近邻矩阵,并结合PPLS模型投影矩阵,提出一种新颖的融合模型信息迁移、鉴别流形正则化以及l_(2,p)-范数约束的BNN建模方法,有效增强BNN模型的鲁棒性、建模精度,同时消除数据的随机不确定性;最后给出迭代优化求解方法获取模型最优参数。在不同规模数据集、不同光照和角度图像数据集对所提算法进行仿真验证,结果表明该算法对不同规模数据集均能取得满意的效果;对图像数据集仿真结果表明其具有很强的鲁棒性和泛化性能。

Broad neural networks(BNN)is considered as a mainstream machine learning method after deep neural networks.However,traditional BNN ignores the stochastic uncertainty and local geometric structure of the data.Therefore,dual model collaborative broad neural network based on probabilistic partial least square(PPLS)and sparse discriminative manifold regularization is proposed to deal with the issues in this paper.In this method,PPLS is firstly performed on the high dimensional samples with original and enhancing features to extract low dimensional latent feature variables,eliminating the uncertainty and redundant features of the data.Based on the local and non-local adjacent matrices constructed by sparse representation and PPLS projection matrix,a novel BNN modeling method is then developed by integrating model information transfer,discriminative manifold regularization l_(2,p)-norm constraint,effectively enhancing robustness and accuracy of the proposed model,and the stochastic uncertainties of the training data are also considered.Finally,an iterative optimized method is developed to yield optimal model parameters.Simulations are performed on various scale datasets and image sets under different types of noises.The results show that the algorithm can achieve satisfactory results for data sets of different scales,and the simulation results for image data sets show that it has strong robustness and generalization performance.