Recognizing emotions from brain signals could support wearable health monitoring and human-computer interaction, but running the necessary models efficiently on portable hardware remains a challenge. Researchers at Beijing Institute of Technology have developed a brain-inspired computing architecture that processes sparse events extracted from electroencephalography (EEG), allowing inactive processing units to pause between events.
The study, published in National Science Review, combines a Dynamic Adaptive Spiking Neural Network (DA-SNN) with a globally asynchronous locally synchronous (GALS) accelerator. The design links feature selection, spike encoding and hardware execution so that sparse activity can translate into lower energy use during inference.
EEG records electrical activity from the brain as continuous signals. Yet not every sampling instant contributes equally to recognizing an emotional state. Conventional synchronous hardware follows a fixed clock and can incur computation and data-transfer costs even when useful activity is sparse. Reducing model size or numerical precision alone does not necessarily remove this overhead.
The new model first calculates power spectral entropy within EEG time windows and maps values from individual electrodes onto a two-dimensional grid. This represents how signal patterns vary across brain regions, frequencies and time. A gating module then selects useful features and suppresses redundant information before the features are converted into spikes.
The encoding method translates stronger feature activations into earlier spikes without floating-point division. An adaptive temporal spiking network then classifies the events, with each neuron emitting at most one spike per inference. Its time windows adjust to the timing of incoming spikes, concentrating computation within informative periods. The measured average firing rate was approximately 35%, and the computational cost was about half that of dense computation.
The hardware uses two local clock domains: one handles tensor computation and spike encoding, while the other performs spiking-network inference. Events move between the domains through a request-and-acknowledgment handshake. Within the spiking array, processing elements activate when events arrive, while idle modules remain inactive. Locally stored weights and reused intermediate events further reduce data movement.
On a Xilinx Zynq-7020 field-programmable gate array (FPGA) operating at 100 MHz, the accelerator completed one inference in 18 microseconds, with total power consumption of 148 milliwatts and energy use of 2.66 microjoules per inference. Compared with EEGNet on the same platform, DA-SNN operated at comparable power while reducing latency from milliseconds to microseconds and using approximately one-fortieth as much energy per inference. Event-driven asynchronous execution further reduced energy per inference by about one-fifth relative to synchronous execution.
The researchers evaluated DA-SNN on five public datasets: SEED, SEED-IV, SEED-V, DEAP and DREAMER. It achieved higher accuracy than more than ten representative comparison models, reaching 96.85% on SEED. Its performance declined less than that of the comparison models under Gaussian noise, low-frequency drift and simulated muscle-signal interference. Reducing weight precision to 8 bits left recognition accuracy nearly unchanged.
The results show how coordinating feature selection, single-spike encoding and event-triggered hardware can improve the efficiency of EEG emotion recognition. The reported hardware measurements concern inference; continuous wearable operation will also depend on bringing signal acquisition and preprocessing online. The architecture provides a path toward low-power affective brain-computer interfaces and continuous emotion sensing.
Jingxin Liu and Yiming Zhang are co-first authors. Qunxi Dong, Lixian Zhu and Bin Hu are co-corresponding authors. All are affiliated with Beijing Institute of Technology. The research was supported by the National Natural Science Foundation of China's National Major Research Instrument Development Project and the Brain Science and Brain-like Intelligence Technology–National Science and Technology Major Project.
National Science Review
Experimental study