Leveraging unlabelled data for generalizable neural population decoding

TL;DR — MOJO is a training framework for spike-tokenizing neural decoders that jointly combines self-supervised masked autoencoding with supervised objectives, letting models exploit unlabelled neural data and outperform purely supervised training—especially when labels are scarce.

• Introduces joint SSL (masked autoencoding) + SL training for spike-level neural data, breaking the prior spike-based models' dependence on paired behavioural labels.
• Validated on three spiking datasets (monkey motor cortex reaching; multi-region mouse vision/decision tasks), with the largest gains in limited-label and few-shot finetuning on new sessions.
• Adding SSL yields more interpretable representations, improving brain-region classification and spike-statistics prediction without explicit optimization for those tasks.
• Generalizes across modalities/species to human ECoG speech, matching purpose-built neuro-foundation models and suggesting more scalable data usage for training NFMs.