Neural Noise- A low-latency neural inference framework enables real-time handwriting recognition from EEG signals on an edge device, supporting portable EEG-BCI applications.1
- Edge deployment enables portable BCI for communication without lab-bound hardware.1 1
Weekly enrichment (2026-07-20)
- The work was published in Scientific Reports (2025, DOI 10.1038/s41598-025-24972-y; also posted as arXiv:2510.19832) and is presented as the first real-time decoding of imagined handwriting from non-invasive EEG deployed fully on a portable edge device.23
- The dataset comprised 32-channel EEG recorded from 15 participants performing imagined handwriting, with preprocessing via bandpass filtering and artifact subspace reconstruction (ASR).3
- The pipeline extracts 85 features spanning time-, frequency-, and graphical-domain descriptors, and the proposed model EEdGeNet is a hybrid architecture combining a Temporal Convolutional Network (TCN) with a multilayer perceptron (MLP).23
- Using the full 85-feature set, EEdGeNet reached 89.83% ± 0.19% character-classification accuracy across the 15 participants, versus 73.82% for an MLP-only and 72.62% for a TCN-only baseline on the same features.2
- Deployed on an NVIDIA Jetson TX2 (256-core Pascal GPU, 7.5–15 W power envelope), per-character inference latency was 914.18 ms, compared with 41.25 ms on a local (non-edge) machine.2
- Reducing the feature set to 10 key features selected by Pearson correlation kept accuracy at 88.84% ± 0.09% (<1% loss) while cutting per-character latency 4.51× to 202.62 ms and lowering per-character inference energy 5.14×.34
- The compact 10-feature set included Hjorth mobility and complexity, RMS, delta-band power, mean dominant frequency, and several graphical ellipse-area descriptors.4
- Unlike cloud/high-performance-compute pipelines or invasive ECoG approaches, this fully on-device non-invasive design targets low-power, portable assistive communication for motor- and speech-impaired users.23