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

• Multiscale fusion combined with spiking neural networks advances invasive BCI neural signal decoding pipelines.¹
• The method is applicable to intracortical signals and is decoding-focused.¹
• SNNs align with neuromorphic deployment for neural time series.^{1 1}

Weekly enrichment (2026-07-20)

• The paper (Song, Han, Zhang, and Xu; Frontiers in Neuroscience, Volume 19, 2025; also on arXiv as 2410.03533) proposes a Multiscale Fusion enhanced Spiking Neural Network (MFSNN) for stable, energy-efficient decoding of invasive BCI neural signals.²³
• The architecture is bio-inspired by the brain’s parallel visual pathways: it splits multi-channel spike trains into sub-paths, and each sub-encoder combines a linear transformation, a spiking channel-attention module, and a temporal convolutional network (TCN), fusing features via skip connections before a spiking classifier.²⁴
• The channel-attention module uses a bottleneck of spiking convolutional layers with Leaky Integrate-and-Fire (LIF) neurons to weight channels, while the TCN uses causal, dilated convolutions (kernel size 3, dilation rate 2) to capture long-range temporal dependencies.⁴
• Robust cross-day generalization is achieved through mini-batch supervised generalization learning, letting the model adapt to new session data with minimal fine-tuning.²
• It was validated on two invasive BCI paradigms in macaque/rhesus monkeys: a single-hand grasp-and-touch task (128-channel M1 recordings at 30 kHz, collected over eight days) and a center-and-out reach task to eight directions (Churchland et al. dataset, two 96-channel arrays, ~2,000 trials/day).⁴
• All models reached high single-day accuracy (>95%), but in cross-day decoding MFSNN maintained higher accuracy (>80%) than MLP and GRU baselines under matched fine-tuning; performance stabilized when the fine-tuning data ratio reached about 7.8% (roughly 48-96 seconds of recording).⁴
• The authors report the MFSNN reduces computational/energy cost by about 90.9% compared with ANN-based counterparts, making it well-suited to neuromorphic hardware for online invasive BCI decoding.⁴
• The authors note the results are from macaque motor tasks and that transferability to human neural data and more complex scenarios (e.g., neuroprosthetics, affective computing) remains untested; they propose integrating neural-manifold learning as future work.⁴

Footnotes

1. https://news.google.com/rss/articles/CBMilAFBVV95cUxNX191a0JwczNWUlYwRUhyTFFjVWEwODdzbl9qSjhLZE9HcXpEVzdKUGNjajkwMjBsaVoydUtxMlZZeVp3MmpJRDAzc3BrY3A2dlh3SFB0R1JkMTVOOWRjR1JNd0RaSnJ5ZmdLZnEyejd3WGRmZXI5bFVNQ0YxX3BPQXpKOVBNRkxkVmdxOVpHRnNSS01f?oc=5 ↩ ↩² ↩³ ↩⁴

2. https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2025.1551656/full ↩ ↩² ↩³

3. https://arxiv.org/abs/2410.03533 ↩

4. https://pmc.ncbi.nlm.nih.gov/articles/PMC11885244/ ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶