Reference Architecture

BCI research has long been slowed by inconsistent data formats, divergent decoder implementations and incompatible deployment toolchains. BCIJelly standardizes 18 BCI datasets into inputs ready for AI training and integrates 15 benchmark decoders and 80 reusable modules. Its automated architecture search generates task-specific decoders without manual design and can extend into a large language model-driven closed-loop mode that supports single-task, multitask and cross-species decoder design; the system also offers interactive visualization software that requires no coding. A single-command pipeline compiles trained decoders onto neuromorphic hardware, cutting power consumption 30- to 50-fold while maintaining decoding performance. The work has been validated in humans, macaques and mice across five paradigms: motor, visual, speech, emotion and auditory. It is a preprint that has not been peer reviewed.

Two national standards for brain-computer interfaces took effect in China on 1 August 2026, the first foundational standards the field has had there. GB/T 47023 defines a reference architecture for BCI systems; GB/T 47127 sets a common format for multimodal data, down to directory structure and file naming. Both carry the GB/T prefix, marking them as recommended rather than mandatory, so they bind no one by law and work only as far as industry and procurement choose to adopt them.