Benefits of a high-performance computing cluster for calibrating brain-computer interface technology

Background: Brain-computer interface (BCI) technology provides an opportunity for the user to control an electronic device using information encoded in electroencephalography (EEG) without involving neuromuscular pathways. However, training AI to accurately decode voluntary modulations in EEG is a challenge and requires significant computational resources.
Example study: Currently, we are running 5 studies with similar datasets. In our 3-dimension (3D) arm movement decoding study, 10 subjects are trained over 10 sessions to achieve real-time control on a virtual arm directly through brain activity. Each session
involves 4 runs with actual movement and 4 runs with imagined movements, 2 blocks per run, and 4 trials per target are involved in each block. This scenario for each session results in 32 trials per target for both actual and imagined movements (256 trials per session in total, 2,560 trials for each subject).
BCI framework: The (artificial intelligence) AI includes extracting time-frequency patterns (MATLAB/Simulink) and translating these to control signals through a convolutional neural network (CNN) and a long-short term memory (LSTM) based kinematic data decoder module (Python). The feedback is presented with the Unity3D game engine.
BCI calibration: The subject-specific calibration of the BCI aims to find hyperparameters that dictate deep neural network architectures and learning ability to enable decoding the kinematic information with maximal accuracy from EEG. The parameters being optimized during the calibration phase can be separated into the following groups: (1) parameters of the preprocessing module for generating optimal features; (2) structural elements of the DL module; (3) Learning parameters of the DL algorithm; (+) Weights within the neural network (NN). Tuning these parameters is a time-consuming process which requires running multiple times the DL algorithm to find a combination of the hyperparameters which results in maximal decoding accuracy (DA).
Calibration time: The training of the DL module using a single instance of hyperparameter set with a high-end PC takes 20 minutes per CV fold for each session. As in our virtual arm study, all variations of the hyperparameter options within the above-described parameter groups count 41,472 options, the calculation time with a grid search would take approximately 13,824 hours (i.e., 576 days) per CV fold for each session. Human-supervised heuristic search can decrease the calculation time where a set of hyperparameter combinations are defined based on results obtained in the previous phase of the search. With this method, the number of calibration steps can be reduced significantly, but this strategy requires human supervision and may find only a local optimum in the full parameter space. Asynchronous successive halving algorithm (ASHA) and RayTune (a Scalable Hyperparameter Tuning) methods reduce the number of calibration steps by using an automatic method for inherently dependent parallelization.
High-performance computing (HPC): HPC facilities such as the Kelvin 2 cluster in the Northern Ireland HPC not only provide faster computation speed but also enable the use of multiple CPU and GPU (Tesla V100 and A100) to parallelize the analysis of the hyperparameter options. Running ASHA/RayTune optimization on the HPC cluster, the single session calibration time can be reduced significantly e.g., using 12 GPU units in parallel for 42 times repeated ASHA/RayTune hyperparameter optimization loop, the computation time from 576 days (full grid search) decreases to 14 hours.
Conclusion: NIHPC reduces the BCI calibration time to comprehensively assess hyperparameters between daily subject training sessions and enables us to perform high-quality research across multiple neurotechnology studies.