A constrained alternating least squares nonnegative matrix factorization algorithm enhances task-related neuronal activity detection from single subject´s fMRI data

This paper proposes a constrained alternating least squares nonnegative matrix factorization algorithm (cALSNMF) to enhance alternating least squares non-negative matrix factorization (ALSNMF) in detecting task-related neuronal activity from single subject´s fMRI data. In cALSNMF, a new cost function is defined in consideration of the uncorrelation and overdeter-mined problems of fMRI data, A novel training procedure is generated by combining optimal brain surgeon (OBS) algorithm in weight updating process, which considers the interaction among voxels. The experiments on both simulated data and fMRI data show that cALSNMF fits data better without any prior information and works more adaptively than original ALSNMF on detecting task-related neuronal activity.