GRAFT Transformer achieves SOTA neural decoding with 9% parameter updates

Neural population activity models are crucial for decoding brain signals in BCIs, but they are typically tied to fixed neuron sets, limiting reuse when recorded neuron identities, counts, or response statistics change over days. GRAFT (Gain-Recalibrated Adapters for Transformer), introduced by Xiangsheng Ge and Yang Xie, solves this by separating reusable temporal dynamics from a recalibratable neuron interface. The neuron interface controls how recorded neurons enter and leave the shared Transformer backbone, with auxiliary gain and positional mechanisms supporting activity modeling. This design allows the model to adapt to new neuron sets by updating only 9.21% of its parameters.

On the standard NLB'21 MC Maze benchmark, GRAFT achieves 0.3866 co-bps (ensemble), setting a new state of the art on the primary co-bps metric. For cross-day recalibration, it reaches 0.3749 co-bps on the MC Maze Large dataset, 0.3112 on Medium, and 0.3152 on Small with restricted target-day support sets. These results demonstrate that the interface-backbone separation enables both strong Transformer-based modeling and data-efficient recalibration. The work is a significant step toward practical, long-term BCIs that work across sessions without full retraining, addressing a key bottleneck in neural prosthetics and brain-machine interfaces.