Finally, beyond spikes, MOJO generalises to human ECoG for speech decoding. It improves over supervised POYO on syllable, consonant, and vowel prediction, and is comparable to foundation models designed specifically for iEEG signals. 💬 5/

With MOJO, we also jointly pretrained on monkey and mouse data – improving mouse decoding and unit-embedding interpretability while preserving performance on monkey reaching tasks. 🐵🐭 4/

MOJO’s unit embeddings are more interpretable: simple linear probes accurately recover region identity and predict various neuronal firing statistics. Their geometry also reflects anatomical proximity and functional relationships. 🧪🔍 3/

Across tasks and species, MOJO improves over supervised counterparts and baselines. Gains are largest with few labels: in a held-out monkey session, adding unlabelled data over 2 and 4 labelled trials yields >60% and >75% of full-supervision performance. 📈 2/

What if spike-tokenising decoders could learn from unlabelled neural data? 🧠⚡ Introducing MOJO, a joint SSL+SL framework for POYO- & POSSM-style models. MOJO improves decoding & few-shot transfer and yields more interpretable unit embeddings across neural datasets & tasks. 1/🧵