The Hyperalignment Method Based On Gradient And Manifold Learning And Its Application

Currently,using both neuroscience and machine learning methods to study how the human brain works is becoming more and more popular.As an imaging technique,functional magnetic resonance imaging has made significant progress in the field of human brain research.It generates image data by measuring neural activity,which not only helps people understand the function of each region of the human brain,but also reveals how the information is encoded in each region.Significant progress has been made in the development of approaches for decoding human neural activities during these years.However,the functional alignment of multi-subject fMRI data remains one of the long-standing important challenges.The typical characteristics of these data are large samples,high dimensionality and non-linearity.In addition,the inherent anatomical and functional variability across subjects make it necessary to do both anatomical and functional alignment before classification analysis.As a result,proper functional alignment methods and good feature selection algorithms are effective ways to solve these problems.In this thesis,two different functional alignment methods are proposed in order to solve the above problems.The main contributions can be summarized as follows:1.For the problems of large samples and high dimensionality in multi-subject fMRI data analysis,this thesis proposes a novel gradient based hyperalignment method called Gradient Hyperalignment.Hyperalignment is currently one of the most effective functional alignment methods and can greatly improve the classification accuracy.Unlike other hyperalignment methods that use Canonical Correlation Analysis to get the solution,Gradient Hyperalignment employs Independent Component Analysis to solve the hyperalignment problem.Actually,the ICA algorithm itself is a kind of effective feature selection algorithm and improves classification accuracy and reduces the runtime by selecting independent features,which are exactly the advantages of Gradient Hyperalignment algorithm.In addition,due to the problem of large samples and high dimensionality of multi-subject fMRI data,the traditional functional alignment method usually generates high runtime and takes up too much memory.As a result,the stochastic gradient ascent algorithm is used for optimizing.Combining the two methods,ICA and the stochastic gradient ascent algorithm,Gradient Hyperlignment algorithm has a lower runtime on the large dataset compared with other algorithms.The empirical studies show that the proposed method can greatly reduce the runtime and improve the classification accuracy.2.For the problems of non-linearity and high-dimensionality in multi-subject fMRI data analysis,this thesis proposes a novel manifold-based hyperlignment method called Mainfold-Based Hyperalignment.Due to the problem of non-linearity and high-dimensionality of multi-subject fMRI data,traditional linear hyperlignment method has big problems in nonlinear data.The kernel hyperlignment method utilizes kernel method to deal with nonlinear problems,but this method ignores the inherent distribution characteristics of the data.Therefore,manifold learning is used to reduce the dimensionality of high-dimensional data.Since most fMRI data processed are related to neural vision,theoretically these data conform to manifold characteristics,so Mainfold-Based Hyperalignment can transform high-dimensional data into low-dimensional manifold data and improve classification accuracy.Like Gradient Hyperlignment,the stochastic gradient ascent algorithm is used for optimization to solve the problem of large samples.Therefore,Mainfold-Based Hyperalignment generates low runtime on big datasets.The empirical studies show that the proposed method can greatly reduce the runtime and improve the classification accuracy.3.To make the research easily reproducible,a GUI-based toolbox was created for running the popular task-based fMRI images analysis algorithms,including the proposed methods in this thesis.