Multiscale fusion enhanced spiking neural network for invasive BCI neural signal decoding

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Neuroscience

Background:

• Brain-computer interfaces (BCIs) require stable, long-term neural signal decoding.
• Spiking Neural Networks (SNNs) are suitable for neural signal processing due to their dynamics.

Purpose of the Study:

• To present a novel Multiscale Fusion enhanced Spiking Neural Network (MFSNN) for efficient and energy-conserving neural signal decoding.
• To improve the accuracy and robustness of BCIs using SNNs.

Main Methods:

• Developed a Multiscale Fusion enhanced Spiking Neural Network (MFSNN) inspired by human visual perception.
• Employed temporal convolutional networks and channel attention for feature extraction.
• Integrated features using skip connections and utilized mini-batch supervised generalization learning for cross-day decoding.

Main Results:

• The MFSNN demonstrated superior accuracy and computational efficiency compared to traditional methods (MLP, GRU) in benchmark invasive BCI tasks.
• Achieved real-time, efficient, and energy-conserving neural signal decoding.
• Showcased improved generalizability and robustness for cross-day signal decoding.

Conclusions:

• The MFSNN offers a significant advancement in neural signal decoding for BCIs.
• The MFSNN's architecture is well-suited for implementation on neuromorphic chips, enabling energy-efficient online decoding.