Neuroscience-Inspired Deep Learning Brain-Machine Interface paper posted

A paper describes a new decoder for brain machine interfaces that aims to decode motor intentions from neural activity. The Single-Direction CNN-LSTM decoder separately models extension and flexion dynamics through parallel CNN-LSTM branches. This design draws from motor cortex encoding mechanisms to address limitations in existing monolithic architectures. Each branch extracts spatial temporal features from neural spike data and predicts directional joint variables. The predictions are then combined by subtraction to produce the net angular velocity and torque of upper-limb joints. The model was tested with invasive recordings from a macaque during a two-dimensional center-out reaching task. When trained on all tasks the decoder achieved performance comparable to a conventional CNN-LSTM. It significantly outperformed both CNN-LSTM and linear regression baselines in cross-target generalization scenarios. The model captured physiologically meaningful co-contraction patterns that offer insights into motor control. These outcomes suggest that incorporating neuroscience-inspired modular decoding into deep neural architectures can enhance adaptability across tasks. The work points to potential applications in prosthetics and rehabilitation.