A Novel Yi-Jing Inspired Dynamic Routing Mechanism for Functional Ultrasound Signal Completion

A Novel Yi-Jing Inspired Dynamic Routing Mechanism for Functional Ultrasound Signal Completion Enshuo Yang EEG_Completion_AI Project February 28, 2026 Abstract This paper proposes a dynamic routing mechanism inspired by the ancient Chinese Yi-Jing eight-trigram philosophy to address noise, sparsity, and missing distortions in functional ultrasound brain-computer interface signals. The mechanism extracts five key features from input signals and dynamically routes them to eight specialized processing paths, each employing distinct activation functions and adaptive strategies. Real GPU-based experiments demonstrate that the Bagua model achieves MSE values of 0.143 to 0.184 and PSNR values of 7.99 to 9.09 dB across the three distortion types, with average MSE reduction of 36.10 percent and PSNR improvement of 47.71 percent compared to traditional baselines. The current team is exploring the extension of this mechanism to large language models in GGUF format to further validate its generalizability. Keywords: Yi-Jing, Bagua Theory, Dynamic Routing, Functional Ultrasound, Brain-Computer Interface, Adaptive Computation, Signal Completion

Introduction

Functional ultrasound offers promising non-invasive access to brain activity, but raw signals often suffer from noise, sparse sampling, and missing segments. Standard models apply fixed processing, limiting their ability to adapt to varying distortion patterns. Inspired by the eight fundamental trigrams in Yi-Jing, this paper proposes a dynamic routing mechanism that classifies inputs and routes them to specialized processing paths accordingly.

Methodology

The model uses a lightweight feature extractor to compute five signal characteristics: noise level, sparsity ratio, missing ratio, peak density, and periodicity score. A simple classifier routes the input to one of eight processing paths, each with a different activation function. The final output combines the selected path with residual connection. Inference is optimized using sequential slicing and early exit, reducing overhead to 1.06 times baseline.

Experiments and Results

Experiments use 9,000 real fUS samples from the NERF dataset with controlled distortions. All models share identical training settings for fair comparison. Real GPU results show:

Noise: MSE 0.143, PSNR 9.09 dB Missing: MSE 0.184, PSNR 7.99 dB Sparse: MSE 0.161, PSNR 8.58 dB

Average improvement: MSE decreased by 36.10 percent, PSNR increased by 47.71 percent.

Discussion

The routing mechanism provides adaptive processing and clear interpretability through trigram mapping. Current efforts focus on extending it to large language models in GGUF format.

Conclusion

This study demonstrates a Yi-Jing-inspired routing approach that significantly improves fUS signal completion. The method offers a new perspective on interpretable and adaptive neural architectures. References The NERF whole-brain fUS dataset. DOI: 10.5281/zenodo.4905862. Fedus, W., et al. Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity. Journal of Machine Learning Research, 2022. Shazeer, N., et al. Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer. arXiv:1701.06538, 2017.