MOJO: masked autoencoding for label-scarce neural decoding

MOJO adds masked-autoencoder self-supervision to supervised neural decoding, helping spike-tokenizing models use unlabelled data when labels are scarce.

TL;DR — MOJO is a joint training framework for spike-tokenizing neural decoders. The abstract says it combines masked-autoencoding self-supervision with supervised objectives, outperforming purely supervised models on three spiking datasets, especially with limited labels and few-shot new-session finetuning. It also reports improved representation probes and a human ECoG speech extension, but provides almost no numerical detail beyond the count of spiking datasets.

Problem: supervised labels constrain neural decoding

The abstract frames the problem as a label bottleneck for neural decoding: "Robust and accurate neural decoders are integral to neurotechnologies such as brain-computer interfaces and closed-loop experiments." It then states that "current spike-based models are restricted to supervised learning (SL), limiting training to datasets with paired behavioural labels."

MOJO targets this bottleneck by letting spike-tokenizing models use self-supervised information from neural recordings in addition to supervised behavioural labels. The article should therefore be read as an abstract-grounded explainer of a training framework, not as a full reproduction recipe.

Full analysis, extracted claims, numerical results, entity graph, FAQ, related work, applications, and BibTeX are available via x402 micropayment.