A brain-inspired computing system developed by researchers at Beijing Institute of Technology classified emotional states from electroencephalography (EEG) data in 18 microseconds, using 2.66 microjoules per inference. The team reports that the design consumed 97.5% less energy than the EEGNet model used for comparison in tests on an FPGA platform.
Described in the journal National Science Review, the system combines a Dynamic Adaptive Spiking Neural Network with a globally asynchronous locally synchronous (GALS) accelerator. It extracts spectral entropy features from EEG signals, converts informative activity into sparse spike events and allows hardware components to remain idle when no events arrive.
The researchers say each artificial neuron fires no more than once during an inference cycle. The network had an average firing rate of about 35%, which they estimate reduces computational demands by half compared with dense networks. The accelerator’s total power consumption was 148 milliwatts.
The system was evaluated on five emotion-recognition datasets: SEED, SEED-IV, SEED-V, DEAP and DREAMER. It reached up to 96.85% accuracy on SEED. The researchers also reported resilience to Gaussian noise, low-frequency signal drift and electromyographic artifacts associated with muscle activity.
The results were measured using a Xilinx Zynq-7020 FPGA. The study presents the architecture as a potential route to lower-power wearable EEG processing, but the reported performance concerns the tested system and datasets.
Sources: Neuroscience News.