MATLAB based tool for extracting hierarchical Sparse Connectivity Patterns (hSCPs) while reducing the effects of covariates such as age, sex and site in multisite datasets. Standard hSCP implementation can be found at https://github.com/DushyantSahoo/hSCP. Implementation of robust hSCPs based on the following references:
- [1] Dushyant Sahoo and Christos Davatzikos. Learning Robust Hierarchical Patterns of Human Brain across Many fMRI Studies. NeurIPS 2021. https://ieeexplore.ieee.org/abstract/document/9285290
- [2] Dushyant Sahoo et al.Generalizable Hierarchical Patterns for identifying MDD patients: A Multisite Study. https://arxiv.org/pdf/2202.11144.pdf
contact: [email protected]
Requirements:
MATLABMATLAB Deep Learning Toolbox
The main algorithm depends on projection algorithm which needs to installed, is described in the upcoming section.
The projection folder contains a Matlab/Mex/C++ implementation of a linear time algorithm for the projection onto the intersection of the L1-ball and L-infinity-ball (box constraint) provided by A. Podosinnikova.
- make sure your Matlab recognizes your gcc compiler
mex -setup C++
- compile the
proj_L1_box.cppfunction
install.m
- see an example of the code's usage
example_projection.m
The current implementation requires a positive semi-definite symmetric matrix, for example, a correlation matrix as an input. Below is the main function and description of input and output.
[W, lambda, error, C, ld, dlnet_predict, dlnet_correct, acc_train] = rhscp_learn(A,k,alpha,loop,eta,beta1,beta2,eps,tol,site,gamma_predict,gamma_cross,mdd,age,sex)
Below are the inputs to the above function
[1] A contains the input matrices in a cell format, for example- A{1}, A{2}, ....
[2] k contains the number of components in each hierarchy, for example, k = [120,20,4], where 120 is the number of nodes in data or the size of the input matrix 20 is the number of components at level 1, 4 is the number of components at level 2. k can also be [120,10] having only one level or [120,40,20,4] having three levels.
[3] alpha is the sparsity level of components at each level; users will have to play with it a little bit to get a balance between sparsity in the components and good approximation error. If k = [120,20,4] then alpha can be [1,1] i.e. sparsity for each level
[4] loop is the number of iterations of gradient descent
[5] eta, beta1, beta2, eps are the hyperparameters for amsgrad
[6] tol is the % change in the error before gradient descent stops
[7] site variable stores the site indices of each subject
[8] gamma_predict is the weightage given to classification loss
[9] gamma_cross is the weightage given to robustness loss
[10] mdd is a binary variable storing information about whether a subject is healthy or patient
[11] age and sex store demograhic information of each subject
There are three outputs- [1] W stores components at different levels in cell format
[2] lambda stores subject specific information
[3] error stores % information captured by the decomposition; ideally it should be decreasing with the iterations.
A test script using simulated data is also provided, which would give the user an idea of the input parameters and how the output looks. Please refer to the papers for more details. I am thankful to Anastasia for providing me with the code for projection operators.
The author welcomes any contribution and also tries to address any bugs or feature requests that may be filed on the issue tracker at https://github.com/DushyantSahoo/rshSCP/issues.
- [1] Sahoo, Dushyant, and Christos Davatzikos. "Learning Robust Hierarchical Patterns of Human Brain across Many fMRI Studies." Advances in Neural Information Processing Systems 34 (2021).
- [2] Sahoo, Dushyant, et al. "Robust Hierarchical Patterns for identifying MDD patients: A Multisite Study." arXiv preprint arXiv:2202.11144 (2022).
- [3] Podosinnikova, Anastasia. Robust Principal Component Analysis as a Nonlinear Eigenproblem. Diss. Universität des Saarlandes Saarbrücken, 2013.
- [4] Reddi, Sashank J., Satyen Kale, and Sanjiv Kumar. "On the convergence of adam and beyond." arXiv preprint arXiv:1904.09237 (2019).