MOJO: Framework for Leveraging Unlabeled Neural Data in Spike-Tokenizing Models

What changed

Researchers have introduced MOJO (Masked autOencoder-based JOint training), a novel framework designed to enhance spike-tokenizing models used in neurotechnologies like brain-computer interfaces. Traditionally, these models have been limited to supervised learning (SL), requiring paired behavioral labels for training. MOJO addresses this by integrating self-supervised learning (SSL) through masked autoencoding alongside SL objectives. This joint training approach allows models to learn from unlabeled neural data, which is abundant in neuroscience research. The framework has been evaluated on multiple spiking datasets from monkey motor cortex during reaching tasks and mouse recordings during vision and decision-making tasks. Results indicate that MOJO outperforms purely SL-trained models, particularly in few-shot fine-tuning scenarios where labeled data is scarce. Furthermore, MOJO demonstrates generalization capabilities beyond spiking data, showing comparable performance to specialized neuro-foundation models (NFMs) on human electrocorticography data during speech tasks. The framework also yields more interpretable neuronal representations, improving performance on tasks like brain region classification and spike-statistics prediction without explicit optimization for these objectives.

Why it matters for builders

MOJO offers a significant advancement for AI builders working with neural data. By enabling the effective utilization of unlabeled datasets, it reduces the dependency on costly and time-consuming data labeling processes. This opens up possibilities for training more powerful and generalizable neural decoders, crucial for developing advanced neurotechnologies. The framework's ability to perform cross-species and cross-modal transfer learning also means that insights and models trained on one type of neural data can be adapted for others, accelerating research and development across diverse neuroscience applications.

Practical impact

AI builders can explore integrating MOJO into their existing spike-tokenizing model pipelines. The framework's joint SSL-SL approach provides a concrete method for augmenting training datasets with unlabeled neural recordings, potentially leading to improved decoding accuracy and robustness, especially when dealing with limited labeled data. Developers working on brain-computer interfaces, closed-loop experiments, or any application requiring the interpretation of neural population activity can benefit from MOJO's enhanced learning capabilities. The demonstrated generalization to human electrocorticography data suggests potential applications in speech decoding and other clinical neurotechnologies. Further investigation into the framework's performance on specific tasks and species relevant to their projects is recommended.

Caveats and source limits

The research presented is a preprint, and the findings are based on evaluations conducted by the authors on specific datasets. Independent verification of MOJO's performance across a wider range of datasets and tasks would strengthen these conclusions. While the paper demonstrates generalization to human electrocorticography, further research is needed to fully understand its efficacy across all neural modalities. The specific implementation details and availability of the MOJO framework's code are not detailed in the provided excerpt, which may limit immediate adoption. The long-term scalability and computational requirements for training MOJO on very large unlabeled datasets are also areas for further exploration.