Efficient Neural Decoder: Mixture-Regularized Bidirectional GRU with Attention
Xiaole Zhang, Jingchen Zuo, Xiecheng Shao · 2024
Mixture-regularized bidirectional gated recurrent unit with attention (BiGAR) boosts the efficiency of decoding brain signals into hand movement trajectories. The novel neural decoder achieves an R-squared of over 0.8 in less than 0.2 ms of computation time on the MC_Maze dataset using fewer than 500 training trials. The expectation maximization (EM) algorithm used to extract neural hidden states improves R-squared and retains relatively low computation time for BiGAR. We further research on how different mixture regularizers impact the model performance. We generate mixture regularizers through pairwise weighted sum mixing of five individual regularizers associated with the Gaussian, Cauchy, Laplace, Sinc-squared, and Sin-fourth probability density functions. Experiments indicate that the improvement of model R-squared with mixture regularizers exceeds that of traditional individual regularizers and no regularizer.