# Can Unlabelled Neural Data Train Better BCI Decoders?

A new training framework called MOJO — short for **Masked autOencoder-based JOint training** — enables spike-tokenizing neural decoders to learn from unlabelled recordings, directly addressing one of the most persistent bottlenecks in clinical [brain-computer interface](https://bciintel.com/glossary/brain-computer-interface) development: the scarcity of behaviorally annotated neural data.

Presented in a preprint posted July 16, 2026 (arXiv:2607.14086), researchers Ximeng Mao, Nanda H. Krishna, Avery Hee-Woon Ryoo, Matthew G. Perich, and Guillaume Lajoie demonstrate that combining self-supervised learning (SSL) via masked autoencoding with conventional supervised learning (SL) objectives outperforms purely supervised models across three spiking datasets. Those datasets span monkey motor cortex recordings during reaching tasks and multi-regional mouse recordings during vision and decision-making tasks. MOJO also generalizes to human [electrocorticography](https://bciintel.com/glossary/electrocorticography) (ECoG) during speech, where it performs comparably to neuro-foundation models (NFMs) built specifically for continuous signals — without being designed for that modality. The gains are most pronounced in few-shot finetuning scenarios, where only a small number of labelled samples from a new recording session are available. This is precisely the regime that determines whether a decoder survives the transition from a controlled lab setting to real-world clinical use.

---

## The Core Problem MOJO Targets

Every intracortical and [ECoG](https://bciintel.com/glossary/ecog) BCI system faces the same fundamental data economics problem: behaviorally paired labels are expensive, time-consuming, and often clinically impractical to collect at scale. A participant with tetraplegia cannot perform hundreds of labelled motor trials indefinitely. Research consortia accumulate vast libraries of neural recordings — from Utah array sessions, Neuropixels experiments, and ECoG grids — that are either unlabelled or incompletely annotated.

Current state-of-the-art spike-tokenizing models, which represent a meaningful advance over traditional threshold-crossing decoders, have been restricted to supervised learning pipelines. That means they can only train on the subset of data where a behavioral label exists alongside each neural snippet. The rest of the data — potentially the majority of recorded neural activity — goes unused during model training.

MOJO attacks this directly. By adding a masked autoencoding SSL objective alongside the standard SL objective, the framework allows the model to learn internal representations of neural population dynamics from unlabelled spiking data. The masked autoencoder component reconstructs masked portions of tokenized spike trains, forcing the model to develop generalizable representations of neural structure without requiring behavioral ground truth.

---

## What the Evaluations Show

The paper evaluates MOJO across three distinct experimental contexts grounded in the source text:

1. **Monkey motor cortex, reaching tasks** — where spike-based decoding of continuous kinematic variables is the dominant benchmark in the field
2. **Multi-regional mouse recordings, vision and decision-making tasks** — a more heterogeneous, multi-area setting that tests cross-region generalization
3. **Human ECoG during speech** — a clinically relevant modality that uses continuous local field potentials rather than discrete spikes

Across all three, MOJO outperforms purely SL-trained models. The speech ECoG result deserves particular attention: MOJO was not designed for continuous signal modalities, yet it achieves performance comparable to NFMs that were specifically engineered for them. This suggests the joint SSL+SL training instills representations robust enough to transfer across neural signal types — a meaningful finding for any team building decoders intended to operate across multiple recording modalities.

The few-shot finetuning results carry the most direct clinical weight. When only a small number of labelled trials from a new session are available — simulating the realistic constraints of a BCI calibration session — MOJO's advantage over SL-only baselines is most pronounced. This is the scenario that governs how quickly a patient can achieve functional BCI control after device implantation or after electrode drift degrades a previously calibrated decoder.

Beyond decoding accuracy, the authors report that SSL training improves **brain region classification** and **spike-statistics prediction** without explicit optimization for those tasks. This emergent interpretability is analytically useful: decoders that encode neurobiologically meaningful structure are more likely to generalize across sessions and participants, and more tractable to debug when performance degrades.

---

## Skeptical Read: What MOJO Doesn't Yet Prove

The paper is a preprint and has not undergone peer review. Several important caveats deserve scrutiny before translating these findings into device-development decisions.

**Dataset scale is unspecified in the abstract.** The source text does not report the number of electrodes, recording sessions, or trial counts used in evaluation. Whether MOJO's SSL gains hold with the electrode counts typical of clinical-grade implants — Utah arrays at 96 channels, Precision Neuroscience's Layer 7 cortical interface at higher densities, or Neuralink's N1 implant — is not demonstrated here.

**The human ECoG data is from research participants, not implanted patients.** Speech ECoG datasets in the literature typically come from epilepsy monitoring patients with temporary subdural grids, not from chronically implanted BCI users. Decoder performance in a chronic implant setting, where signal statistics shift over months due to glial scarring and electrode impedance changes, is a different problem.

**Cross-subject generalization is not the same as cross-patient generalization.** The multi-session and multi-species breadth is genuinely impressive, but moving from monkey-to-mouse-to-human research datasets to a clinical trial cohort introduces regulatory and signal-quality variables the paper does not address.

**No bits-per-second or information transfer rate benchmarks are reported.** For the BCI engineering community, raw decoding accuracy improvements need to be translated into throughput metrics to assess functional relevance.

---

## Industry and Clinical Translation Implications

The practical bottleneck MOJO addresses — label scarcity — is real and commercially significant. Companies building high-density intracortical systems like [Blackrock Neurotech](https://bciintel.com/companies/blackrock-neurotech) and [Precision Neuroscience](https://bciintel.com/companies/precision-neuroscience) accumulate large unlabelled neural datasets across their research and clinical deployments. A training framework that extracts value from that unlabelled data has direct implications for reducing the per-patient calibration burden and improving decoder longevity between recalibration sessions.

For [closed-loop BCI](https://bciintel.com/glossary/closed-loop) systems — where decoder drift during chronic use is an ongoing engineering challenge — the few-shot finetuning capability is potentially more valuable than raw single-session accuracy. A model that can rapidly re-adapt to a new session's neural statistics with minimal labeled data reduces the clinical overhead of maintaining a functional BCI system over months or years of use.

The cross-modality generalization to ECoG also matters for companies building less invasive systems. [Synchron](https://bciintel.com/companies/synchron)'s endovascular Stentrode records a signal more analogous to ECoG than intracortical spiking activity. A decoder architecture that handles both modalities within a single framework could simplify the software stack across device classes.

For researchers working at the intersection of neural decoding and robotic motor control — an area with growing interest from groups developing humanoid platforms that accept neural command signals — the motor cortex decoding advances here are directly relevant; [humanoidintel.ai](https://humanoidintel.ai) covers that convergence in depth.

The neuro-foundation model trajectory is also worth watching. MOJO positions itself as complementary to existing NFMs rather than a replacement, and its ability to match NFM performance on ECoG speech decoding without being specifically designed for continuous signals suggests that spike-tokenizing architectures augmented with SSL may close the gap with continuous-signal NFMs faster than the field has assumed.

---

## Key Takeaways

- **MOJO** (Masked autOencoder-based JOint training) combines SSL via masked autoencoding with supervised learning to train spike-tokenizing neural decoders on unlabelled data
- Evaluated across monkey motor cortex, multi-regional mouse recordings, and human ECoG speech datasets — outperforming purely supervised models on all three
- **Few-shot finetuning** gains are the most clinically significant result: MOJO adapts rapidly to new sessions with minimal labelled data
- SSL training produces emergent interpretability improvements in brain region classification and spike-statistics prediction without explicit optimization
- Generalizes to ECoG, a continuous-signal modality, performing comparably to NFMs designed specifically for it
- Preprint only — no peer review, no electrode-count or bits-per-second benchmarks, and no chronic implant validation reported
- Directly addresses a commercial bottleneck: extracting value from the large unlabelled neural recording archives that clinical BCI programs accumulate

---

## Frequently Asked Questions

**What is MOJO in the context of neural decoding?**
MOJO (Masked autOencoder-based JOint training) is a training framework introduced in arXiv:2607.14086 that allows spike-tokenizing neural decoders to learn from unlabelled neural recordings by combining self-supervised masked autoencoding with supervised learning objectives. It was developed by Mao, Krishna, Ryoo, Perich, and Lajoie and evaluated on monkey, mouse, and human ECoG datasets.

**Why does label scarcity matter for BCI decoders?**
Behaviorally labelled neural data — where each spike train snippet is paired with a known motor command or sensory state — is expensive and time-limited to collect, particularly from clinical participants. Most neural recordings go unlabelled, leaving large datasets unused in supervised-only pipelines. MOJO allows those unlabelled recordings to contribute to model training.

**Does MOJO work on human ECoG data?**
Yes. The paper reports that MOJO generalizes to human electrocorticography during speech tasks, achieving performance comparable to neuro-foundation models specifically designed for continuous neural signals, despite MOJO being architected around discrete spike tokenization.

**What is few-shot finetuning in BCI decoding?**
Few-shot finetuning refers to rapidly adapting a pre-trained decoder to a new recording session using only a small number of labelled trials from that session. This is clinically important because BCI users cannot always provide extensive labeled calibration data, and neural signal statistics shift over time due to electrode drift and biological changes.

**How does this relate to neuro-foundation models?**
Neuro-foundation models (NFMs) are large pre-trained neural data models analogous to language foundation models. MOJO's SSL augmentation brings spike-tokenizing architectures closer to NFM-level performance on tasks those NFMs were designed for, and the authors suggest this is a path toward more flexible and scalable NFM training that incorporates both labelled and unlabelled data.

---

*This article is based on a preprint (arXiv:2607.14086) that has not yet undergone peer review. All findings are from research-stage experiments and do not constitute clinical guidance. Results from animal and research-participant datasets should not be extrapolated to regulatory conclusions about commercially deployed BCI systems.*