Application And Research Of Brain Cognitive Status Based On A Modified Sparse Tensor Model

The study of Cognitive Neuroscience is to uncover the cognitity principle of thebrain. Functional magnetic resonance imaging is one of the most effective means in thestudy of brain science. The means of Machine study is the most effective ways in thestudy of the brain which can get effective information from the primitive photograph,taken as the basic in exploring the secret of the brain.Brain magnetic resonance data is essentially a four-dimensional tensor data. In theface of high-dimensional data, we typically represent them as one-dimensional vectors,which not only destroys the inner structure and the potential information of the originaldata, but also breaks inconvenience to the subsequent study. In doing so, thetensor-based decomposition method have been proposed to conquer theseshortcomings and been used to the cognitive neuroscience.This paper proposes a new sparse non-negative tensor factorization algorithmwhich is used to analysis and disposal of nuclear magnetic resonance image, acombination of with the characteristics of support vector machine(SVM), successfullyimplements the brain’s visual cognitive states classification. Sparse Non-negativetensor factorization (SNTF) is a kind of large-scale low tensor rank approximationtechnique, especially for the large-scale and high-dimensional data. In virtue of a strictl1norm regularization and non-negative constraint, SNTF can guarantee the numericalcharacteristic factor and the sub-tensors are sparse and non-negative, and retains theessential information of the original tensor data compared with the traditional tensorfactorization methods. Which not only enhances the efficiency of operation, but alsothe original data can be descripted additively and the model is more consistent with thelaw that people understand the world.This thesis focuses on SNTF-based machine learning method. Firstly, the brainfunctional magnetic resonance data need to be preprocessed and be represented as alarge data tensor, simultaneously constructing high order sparse non-negative tensormodel in the tensor level. Secondly, feature extraction is in progress at each dimensionof fMRI cognitive tensor data, in order to achieve the sparse and non-negativecharacteristic sub-tensor. Finally a combination of with the characteristics of SVM,effectively realizes the classification of specific brain visual cognitive state. Theexperimental results show that SNTF algorithm has a higher performance indimensionality reduction level, compared with TUCKER and PARAFAC algorithms. Feature sparsity which is extracted by SNTF performs remarkable comparatively NMFalgorithm. Besides, In the view of classification accuracy, SNTF+SVM method hasoutstanding recognition effect.