GPU-Accelerated Multivariate Empirical Mode Decomposition for Massive Neural Data Processing

This paper presents an efficient implementation of multivariate empirical mode decomposition (MEMD) algorithm, a multivariate extension of EMD algorithm. Analogous to EMD, MEMD decomposes a multivariate signal into its intrinsic mode functions using joint rotational mode. The algorithm is computationally intensive because of its recursive nature and any increase in input data size results in a non-linear increase in its execution time. Therefore, it is extremely time-consuming to obtain a decomposition of signal, such as EEG into its intrinsic modes using MEMD. As the interest in applying MEMD algorithm in various domains is increasing, there is a need to develop an optimized implementation of the algorithm, since it requires repeated execution of the same operations and computationally extensive interpolations on each projected vector. This can be done using GPGPU, because it has the power to process similar function on different blocks of data. We have compared the optimized implementation of MEMD, using GPU, with the MATLAB implementation for hexa-variate and hexa-deca-variate data sets, and observed that the GPU-based optimized implementation results in approximately $6\times \sim 16\times $ performance improvements in terms of time consumption.