CUDAICA: GPU optimization of Infomax-ICA EEG analysis

Area of Science:

• Neuroscience
• Computational Neuroscience
• Bioinformatics

Background:

• Independent Component Analysis (ICA) is a standard technique for identifying underlying data dimensions in neuroscience.
• The computational cost of ICA currently limits its application in real-time scenarios, such as brain-computer interfaces.
• Existing ICA implementations are computationally intensive, hindering online data analysis.

Purpose of the Study:

• To accelerate Independent Component Analysis (ICA) for neuroscience applications.
• To overcome the computational bottlenecks of ICA for real-time data processing.
• To enable the practical use of ICA in brain-computer interfaces.

Main Methods:

• Profiling the ICA algorithm to identify computational bottlenecks, specifically vector-matrix and matrix-matrix multiplications.
• Developing a GPU-based solution leveraging graphics processing units for accelerated computation.
• Optimizing linear algebra operations by implementing a custom GPU-based approach, bypassing standard CUBLAS routines to avoid CUDA kernel launch overhead.

Main Results:

• Achieved a significant 25-fold increase in ICA calculation speed.
• Identified vector-matrix and matrix-matrix multiplications as the primary computational bottlenecks (approx. 80% of processing time).
• Demonstrated superior performance compared to standard BLAS and CUBLAS implementations.

Conclusions:

• A GPU-based ICA implementation offers a substantial performance improvement.
• The developed method makes real-time ICA analysis computationally feasible.
• This advancement has significant implications for the development of advanced brain-computer interfaces.