High-Speed Neural Signal Inferencing for Handwritten Character Recognition on a Portable Hardware Device

Brain-computer interfaces (BCIs) hold immense potential in assisting individuals with severe motor and communication disabilities by enabling neural signal-based activity recognition, such as handwriting. This study presents the very first implementation of neural signal inference on a portable hardware device, facilitating efficient handwritten character recognition on resource-constrained platforms. Neural signals from a publicly available dataset are processed into neural spike-event data, facilitating the classification of 31 handwritten characters on an NVIDIA Jetson TX2. To enhance model generalization and mitigate overfitting, random noise injection and time-shifting-based data augmentation techniques are applied. The proposed approach utilizes EfficientNetB0 with neural spikes, and achieves 99.17% test accuracy, significantly outperforming previous model results. During high-speed inference, EfficientNetB0 achieved a Word Error Rate (WER) of 0.96% and a Character Error Rate (CER) of 0.2%, with a character decoding latency of 37.5 milliseconds on the Jetson TX2 while processing 100 sentences used in daily life. These results validate the feasibility of accurate high-speed neural decoding on portable edge hardware, highlighting the impact of lightweight machine learning models in BCI applications.