Study On Matrix Randomized Autoencoder Algorithm

Randomized autoencoder(RAE)has been widely studied in various fields because of its fast training speed and superior representation learning ability,especially the deep randomized neural network based on stacked randomized autoencoder has shown great potential in the field of anomaly detection.However,the randomized encoder algorithm still has some shortcomings,mainly manifested in: 1)the representation ability of multidimensional tensor data is weak.The existing randomized encoders are based on vector form,that is,the input needs to be vectorized in advance,which seriously breaks the structure information of multidimensional tensor data.2)Easy to be disturbed by random coding problems.Due to the lack of effective regular term to constrain the encoding output,it is easy to result in invalid encoding representation of RAE output.Around these two problems,this paper has carried out the following research work:1.Design of matrix randomized autoencoder for single channel two-dimensional data.In this paper,a one-side matrix randomized autoencoder(MRAE)is proposed for singlechannel two-dimensional data(such as gray level map).It mainly includes two parts:1)OMRAE algorithm,which extracts the row or column structure information of twodimensional data through unilateral direct matrix mapping; 2)Two unilateral matrix randomized autoencoders(DMRAE)are used in parallel to extract data structure information.Finally,stack matrix randomized autoencoder to build deep random neural network.The proposed method is verified on several public standard data sets,demonstrating the strong characterization ability of matrix randomized autoencoder.2.Design of multi-channel matrix randomized autoencoder for multi-dimensional tensor data.MRAE can only effectively represent the two-dimensional data of a single channel.Therefore,this paper further extends it to the representation learning of multi-dimensional tensor data,and proposes a multi-channel matrix randomized autoencoder model(MMRAE).Similarly,a one-side multi-channel matrix randomized encoder and a double-side multi-channel matrix randomized encoder are included.The main improvement is to implement random matrix mapping and fusion for each channel respectively,and then reconstruct the data of each channel from the fusion output with minimum loss.In this process,the representation of multi-channel tensor data is learned.The superiority of the proposed MMRAE is verified by experiments on several image data sets.3.In order to solve the problem of invalid coding of randomized autoencoders,this paper further proposes the constraints of within-class scatter and within-class interaction distance in matrix form.The within-class scatter constraint takes the mean vector of data as the center to measure the dispersion degree of within-class samples.within-class interaction distance is the extension of within-class divergence distance,which can be regarded as the within-class divergence distance centered on each sample respectively,to solve the problem that the divergence distance is too dependent on the mean value.Based on these two constraints,a matrix randomized encoder based on within-class divergence constraint(WSI-MRAE)and a matrix randomized encoder based on within-class interaction distance constraint(WID-MRAE)are constructed respectively,and their results are proved to be superior to the unconstrained model on multiple datasets.