MOJO Uses Unlabelled Neural Data to Improve Generalizable Decoding

Introduction

Neural decoders are a core component of brain-computer interfaces, closed-loop neuroscience experiments, and future neurotechnology systems. Their job is to translate population neural activity into useful variables such as movement, perception, or speech-related information. Yet training these systems usually depends on paired neural and behavioural labels, which are costly to collect and often limited across sessions, subjects, or tasks.

The arXiv paper “Leveraging unlabelled data for generalizable neural population decoding” proposes MOJO, short for Masked autOencoder-based JOint training. The main idea is straightforward but important: spike-tokenizing neural models should not be restricted to purely supervised learning. They can also learn from unlabelled neural activity through a masked autoencoding objective.

Key points

• Using unlabelled neural recordings: Existing spike-level tokenization models have shown strong decoding results, but they generally rely on labelled datasets. MOJO adds a self-supervised route, allowing the model to learn structure from neural activity even when behavioural labels are absent.
• Joint supervised and self-supervised training: The framework combines masked autoencoding with a supervised decoding objective. In practice, this means the model learns both the internal regularities of neural signals and the mapping to behavioural outputs.
• Evaluation across multiple spiking datasets: The authors test MOJO on three spiking datasets, including monkey motor cortex recordings during reaching tasks and multi-region mouse recordings during vision and decision-making tasks.
• Stronger gains when labels are scarce: Compared with purely supervised models, MOJO performs better overall, with the most pronounced improvements in limited-label settings. The paper highlights few-shot finetuning on new sessions as a particularly relevant use case.
• More interpretable representations: Adding self-supervision also improves performance on brain-region classification and spike-statistics prediction, even though these tasks are not explicitly optimized during training.
• Beyond spikes: The authors further apply MOJO to human electrocorticography during speech. There, it still outperforms purely supervised models and reaches performance comparable to neuro-foundation models designed for continuous signals.

Why it matters

MOJO’s significance lies in changing what counts as useful training data for neural decoding. In many neuroscience and clinical settings, unlabelled neural recordings are easier to accumulate than carefully aligned behavioural labels. If models can learn transferable structure from those recordings, the cost of adapting decoders to new sessions or users could be reduced.

The work also fits into the broader movement toward neural foundation models: systems that can be pretrained across richer datasets and then adapted to downstream tasks. MOJO suggests that self-supervised objectives can help spike-tokenizing models become more flexible, more data-efficient, and more robust across tasks, species, and signal modalities.

As an arXiv preprint, the method still needs further validation in more realistic long-term and clinical settings. Still, the paper points to a practical direction for neurotechnology: use the large amount of unlabelled neural data that would otherwise sit outside the supervised training loop.

Source: arXiv