MOJO framework uses unlabelled data to boost neural decoding performance
MOJO trains spike-tokenizing models with self-supervised learning, even without labels.
A team of researchers from multiple institutions (including Ximeng Mao, Nanda H. Krishna, Avery Hee-Woon Ryoo, Matthew G. Perich, and Guillaume Lajoie) introduced MOJO (Masked autOencoder-based JOint training), a training framework designed for spike-tokenizing neural network models. MOJO addresses a key limitation of current spike-based decoders: their reliance on supervised learning, which requires paired behavioral labels. By jointly optimizing a self-supervised masked autoencoding objective alongside a supervised loss, MOJO can leverage large amounts of unlabelled neural data during training.
Evaluated on three diverse spiking datasets—monkey motor cortex during reaching tasks, multi-regional mouse recordings during vision and decision making, and human electrocorticography (ECoG) during speech—MOJO consistently outperformed purely supervised models. The gains were especially pronounced in few-shot finetuning scenarios with minimal labelled data from new sessions. Additionally, MOJO produced more interpretable neuronal representations, improving brain region classification and spike-statistics prediction without being explicitly optimized for those tasks. The framework also generalized to continuous neural signals (ECoG) and achieved performance comparable to dedicated neuro-foundation models. Overall, MOJO offers a path toward more flexible and scalable use of unlabelled data in training neural decoders for brain-computer interfaces and neuroscience research.
- MOJO combines self-supervised masked autoencoding with supervised learning on spike-tokenizing models, allowing use of unlabelled data.
- Outperformed purely supervised models on three datasets (monkey, mouse, human) and excelled in few-shot finetuning with limited labels.
- Improves interpretability (brain region classification, spike-statistics prediction) and generalizes to human ECoG without task-specific tuning.
Why It Matters
Enables robust neural decoders for BCIs using abundant unlabelled data, reducing costly labeled experiments.