Leveraging unlabelled data for generalizable neural population decoding
MOJO adds masked-autoencoder self-supervision to spike-token neural decoders so they can learn from unlabelled neural data.
The paper reports better decoding than purely supervised spike-token models across monkey motor-cortex and mouse vision/decision datasets. The gains are strongest when labelled data is scarce, including few-shot finetuning on a new session. The authors also say the learned neuronal representations improve brain-region classification and spike-statistics prediction without being trained for those targets. They extend the result to human electrocorticography during speech, where MOJO matches continuous-signal neuro-foundation models while outperforming supervised baselines. ArXiv · AI/CL/LG's note
The paper reports better decoding than purely supervised spike-token models across monkey motor-cortex and mouse vision/decision datasets. The gains are strongest when labelled data is scarce, including few-shot finetuning on a new session. The authors also say the learned neuronal representations improve brain-region classification and spike-statistics prediction without being trained for those targets. They extend the result to human electrocorticography during speech, where MOJO matches continuous-signal neuro-foundation models while outperforming supervised baselines. ArXiv · AI/CL/LG's note
score 4