MOJO: Leveraging Unlabeled Data for Generalizable Neural Population Decoding

What changed Researchers have developed a new training framework called MOJO (Masked autOencoder-based JOint training) designed to enhance neural decoders. Traditional spike-based models for neural decoding, such as those used in brain-computer interfaces, have been limited to supervised learning (SL), requiring paired behavioral labels for training. This constraint restricts their applicability when such labeled datasets are unavailable or limited. MOJO addresses this by integrating self-supervised learning (SSL) through masked autoencoding with SL objectives. This joint training approach allows models to leverage unlabeled neural data alongside labeled data. The framework has been evaluated on diverse spiking datasets from monkey motor cortex during reaching tasks and multi-regional mouse recordings during vision and decision-making tasks. Results indicate that MOJO outperforms purely SL-trained models, with particularly notable improvements in few-shot fine-tuning scenarios where only a small amount of labeled data from a new session is available. Furthermore, the incorporation of SSL in MOJO leads to more interpretable neuronal representations, enhancing performance on tasks like brain region classification and spike-statistics prediction without explicit optimization for these specific objectives. The framework's versatility is further demonstrated by its successful application to human electrocorticography data during speech, where it achieves performance comparable to neuro-foundation models (NFMs) specifically designed for continuous signals, outperforming purely SL-trained models.

Why it matters for builders For AI builders, MOJO presents a significant step towards more flexible and scalable data utilization in neural data processing. The ability to effectively use unlabeled data reduces the bottleneck of data annotation, which is often a costly and time-consuming process in neuroscience and neurotechnology development. This framework can empower builders to develop more performant and generalizable neural decoding models, even with limited labeled datasets. The improved interpretability of neuronal representations also offers valuable insights for understanding neural mechanisms, which can inform the design of more sophisticated AI systems.

Practical impact The practical impact of MOJO lies in its potential to accelerate the development and deployment of advanced neurotechnologies. For brain-computer interfaces (BCIs), this means more accurate and responsive control signals, even for users with limited training data. In closed-loop experimental setups, MOJO can enable more sophisticated real-time analysis and intervention based on neural activity. The generalization capabilities of MOJO to different neural modalities, such as electrocorticography, suggest its broad applicability across various research and clinical domains. By enabling the use of unlabelled data across different tasks and species, MOJO paves the way for training more robust and adaptable neuro-foundation models.

Caveats and source limits The primary source for this information is a research paper available on arXiv. While the paper details the methodology and experimental results, it does not provide specific benchmark numbers, release dates for the MOJO framework, or details on its availability as open-source code. The performance improvements are demonstrated on specific datasets and tasks, and further validation across a wider range of neural data types and experimental conditions may be necessary to fully assess its generalizability. The paper also does not include information on potential computational costs or hardware requirements for implementing MOJO. The claims regarding performance are based on the authors' evaluations and comparisons within the context of their study.