Pretraining on 50,000 Hours of Unlabeled Brain Data
How self-supervised encoders kept neural decoders working for weeks without recalibration and set an 11.32 BPS record.
Calibration: a Sisyphean task
Until recently, brain-computer interfaces (BCIs) were an incredible yet inaccessible technology found only in research labs. Their utility relied on bulky hardware and skilled researchers — the furthest thing imaginable from a consumer product. For the last few years, we’ve been evolving BCIs from sci-fi curiosity to an in-home product.
Seeing a user move a cursor for the first time is a surreal experience. Yet despite how incredible direct thought control is, the current user experience falls short. Several UX issues come from the need of a decoder; machine learning models trained to convert neural activity into intended action. Decoders are unique to each user, and all users have to perform calibration prior to using the device. During calibration, users follow a set of structured tasks that generates a training dataset. The accuracy of the resulting model determines the quality of their user experience.
The first calibration is the longest, but it is only the beginning. Neural data is highly nonstationary, and even the best decoders degrade over time. As control worsens and labeled data grows stale, users must repeat calibration to regain performance they once had. This traps our users in a Sisyphean cycle: to use their BCI, they repeatedly perform the same task, generating results that are eventually discarded. Users spend an average of 55 minutes per week calibrating simply to maintain control.
Each unique feature (e.g., typing, gaming) also requires its own decoder. The more full-featured and performant the device, the higher the tax we’ve levied on their time. Calibrating each decoder to peak performance could take the better part of their mornings. The system works, but it asks users to keep earning back control that they had previously achieved.
In the two years since our clinical trial began, our participants have streamed over 50,000 hours of unlabeled, freeform neural data. Our first participant alone has recorded over 9,000 hours — or 22.4 billion spikes. Until now, nearly all of it was unused. Where once we had limited data and relied on supervised techniques, we now have a dataset large enough to pretrain on.
Towards a stable representation
Our goal is to create a universal, drift-robust neural foundation model. Towards that end, we started with participant-specific neural encoders pretrained on thousands of hours of their own neural recordings. These encoders transform the noisy, drift-prone spiking data into stable embeddings useful for decoding. When building these models we placed five bets:
Neural activity is best modeled by a dynamical system1. The activity of a neural population is commonly believed to represent a state evolving under learned rules. This makes the Mamba2 architecture a well-suited inductive bias. Mamba also satisfies our constraint of constant, low-latency inference, further reinforcing the choice.
Treat spiking events like tokens. Inspired by POYO3, each recording channel gets its own learned embedding only seen by the model when it fires. A Perceiver4-based input layer naturally handles variable amounts of activity. This architectural choice also allows models to be easily trained on multiple users.
Exploitable cross-channel structure exists3,5. When designing our training objective, we worked backward from our production and data constraints: the deployed decoder must be fully causal, a user’s future intention is unknowable past a certain horizon, and spiking data is heavily autocorrelated. This led us to choose spatially masked auto-Poisson regression. We feed the model half the channel set and ask it to predict the activity of the other half, forcing it to represent the underlying population state from partial representations.
Intention lives in a stable subspace. This idea has been explored in prior work6,7 and makes intuitive sense. Our pretrained models also seem to support this. As seen below, day-specific information is less dominant in middle layers but re-emerges in the final output.
One neural encoder can serve many tasks. Your hand already does; you can write, play instruments, drive, and more without ever needing a new “attachment”. Since the Neuralink implant ("Link") records from the hand-knob8 region of the motor cortex, the Link should be just as adaptable as your brain and hands.
What changed for our users
Since introducing this feature, users’ decoders are more accurate and long-lived. These improvements are reported through bits per second (BPS), which measures max information transfer rate when using the cursor. BPS weighs both speed and precision. It is the closest thing we have to an objective metric for decoder quality. The median Neuralinker BPS is roughly 10 — find out yours here.
Raising the ceiling
Pretraining on large amounts of unlabeled intracortical data gave us better decoders than we could build from limited labeled data alone. With these decoders, six participants reached personal records. Three surpassed our previous record of 10.39 BPS, and one reached 11.32 BPS — a new BCI record.
Users described the cursor as more responsive, fluid, and steady. They could make precise selections without giving up top speed. The improved control also made several software corrections, including smoothing and low-velocity suppression, unnecessary.
Better than the prior model. Smoother, more directionally accurate, felt effortless. Didn’t have to push or anything, it just glided where I wanted it to go.
— P9
Click decoding similarly improved in both raw performance and overall feel. Decoded probabilities are more confident, rise faster, and distinguish click types more clearly. Users experience fewer accidental clicks — a particularly frustrating source of errors. These gains were key to reaching 11 BPS. Previously, we relied on post-processing to reduce errors at the cost of added delay. Removing that tradeoff makes clicks feel snappy and precise.
[Control] just feels better. [Click behavior] is better for games, quick and responsive. Drop sensitivity a tiny bit and I could easily hit 11.
— P15, right before achieving 11 BPS
Longevity and data efficiency
Just as important is how long a decoder maintains performance without recalibration. With the old system, users typically spent 10 minutes calibrating at the start of each day. Some waited a few days between calibrations, trading performance for time. Few decoders lasted longer than that.
Pretraining has substantially extended decoder lifespan. Even users who recalibrate at the first sign of declining performance went more than a week without wanting to. One user reached 10 BPS with the same decoder on five consecutive days. Some decoders maintained strong performance for over three weeks of use; in the case of one study participant, a decoder remained controllable more than a year and a half later. These early results hint at a future where decoders never degrade, bringing us closer to a BCI that users can simply turn on and use.
In addition to calibrating less often, we’ve found users need less time within calibration to return to peak performance. Some users have had their calibration burden shrink from 10 minutes a day to 10 minutes a week. As seen in Figure 4, 30 seconds of labeled embeddings is equivalent to ~3.5 min of raw spikes and generalizes to data a month out.
The gains in longevity extended to our robotic arm simulation, where users control complex movement in seven or more dimensions simultaneously. Decoders for this more complex task tend to degrade faster than cursor decoders: a strong decoder can become unusable in just six days, forcing users to spend more time recalibrating. Decoders trained on learned embeddings remained controllable a week later, allowing users to skip the early stages of calibration.
Cursor teleportation
It’s well established that motor cortices encode information beyond muscle activations and velocity intentions. Prior works have explored decoding position9, error signals10, and preparatory intention11 to improve BCI systems. We’re currently exploring decoding the screen coordinates of the cursor and target purely from neural activity. Figure 7 below visualizes some preliminary results; predictions are from stateless linear models fit purely on embeddings.
These embeddings also let us explore forms of control beyond what a mouse can offer. A mouse reports movement and clicks; neural signals may also reveal when a user is about to act or realizes they’ve made a mistake. What would an interface designed around the brain really be capable of? One possibility is cursor teleportation: predicting where a user wants to go and moving the cursor there in a single step.
Figure 8 above shows this idea in action. We can’t yet predict the intended target with pixel-perfect precision, but the system substantially reduces cursor travel time and crossed 10 BPS in initial testing. With further refinement, we can deliver a completely new class of computer interaction experience.
Future directions
Pooling brains. All results from live use in this post came from models pretrained on a single participant’s data. The next step is to pool data across participants and test whether a shared foundation model improves with scale. We have yet to improve performance online: decoders built from multi-participant models perform no better than their single-participant counterparts. We have, however, seen successful transfer between participants.
In one experiment, a model trained exclusively on P9’s data successfully transferred to P2 despite keeping 99% of the weights completely frozen. Even better, this model outperformed P2’s pre-existing decoder. Other work such as POYO3 and NDT25 have also demonstrated (offline) wins from pooling individuals together, further supporting that such a model is feasible. A successful pooled model could give new users a stronger starting point, simplify onboarding, and let each participant’s data improve control for everyone else.
One-shot calibration. Could a user calibrate just once? Our recent results give us reason to pursue that goal. Maintaining performance for a week is encouraging; extending it to a year or more would require decoders that adapt as neural signals change. Usable control 20 months after calibration suggests that useful information can persist. The challenge is to keep decoding it reliably, so users can spend less time maintaining their BCI and more time using it.
Zero-shot calibration. Could someone achieve able-bodied control without ever calibrating? A model that generalizes to all humans could make control possible out of the box. The goal is simple: wake up from surgery with a device that matches your intent.
An API for the motor cortex. We use the same hands to type, play games, and pick up a glass. Could one decoder rule them all? We envision a shared decoder for all intended hand movement: an API for the motor cortex. Users could calibrate once across applications, and developers could build on a common interface. This would make it easy for developers to build applications for the Link and pave the way for a rich BCI app ecosystem.
None of these problems are solved. Every solution has the potential to improve lives and teaches us about the brain. If any of this excites you, please reach out — we’re hiring ML engineers.
References
- Pandarinath, C., O’Shea, D. J., Collins, J., Jozefowicz, R., Stavisky, S. D., Kao, J. C., Trautmann, E. M., Kaufman, M. T., Ryu, S. I., Hochberg, L. R., Henderson, J. M., Shenoy, K. V., Abbott, L. F. & Sussillo, D. Inferring single-trial neural population dynamics using sequential auto-encoders. Nature Methods 15, 805–815 (2018).
- Gu, A. & Dao, T. Mamba: Linear-Time Sequence Modeling with Selective State Spaces. arXiv 2312.00752 (2023).
- Azabou, M., Arora, V., Ganesh, V., Mao, X., Nachimuthu, S., Mendelson, M. J., Richards, B., Perich, M. G., Lajoie, G. & Dyer, E. L. A unified, scalable framework for neural population decoding. Advances in Neural Information Processing Systems 36 (2023).
- Jaegle, A., Gimeno, F., Brock, A., Vinyals, O., Zisserman, A. & Carreira, J. Perceiver: general perception with iterative attention. Proceedings of the 38th International Conference on Machine Learning 139, 4651–4664 (2021).
- Ye, J., Collinger, J. L., Wehbe, L. & Gaunt, R. Neural Data Transformer 2: multi-context pretraining for neural spiking activity. Advances in Neural Information Processing Systems 36 (2023).
- Karpowicz, B. M., Ali, Y. H., Wimalasena, L. N., Sedler, A. R., Keshtkaran, M. R., Bodkin, K., Ma, X., Rubin, D. B., Williams, Z. M., Cash, S. S., Hochberg, L. R., Miller, L. E. & Pandarinath, C. Stabilizing brain-computer interfaces through alignment of latent dynamics. Nature Communications 16, 4283 (2025).
- Gallego, J. A., Perich, M. G., Chowdhury, R. H., Solla, S. A. & Miller, L. E. Long-term stability of cortical population dynamics underlying consistent behavior. Nature Neuroscience 23, 260–270 (2020).
- Willett, F. R., Deo, D. R., Avansino, D. T., Rezaii, P., Hochberg, L. R., Henderson, J. M. & Shenoy, K. V. Hand knob area of premotor cortex represents the whole body in a compositional way. Cell 181, 396–409.e26 (2020).
- Wang, W., Chan, S. S., Heldman, D. A. & Moran, D. W. Motor cortical representation of position and velocity during reaching. Journal of Neurophysiology 97, 4258–4270 (2007).
- Even-Chen, N., Stavisky, S. D., Kao, J. C., Ryu, S. I. & Shenoy, K. V. Augmenting intracortical brain-machine interface with neurally driven error detectors. Journal of Neural Engineering 14, 066007 (2017).
- Kaufman, M. T., Churchland, M. M., Ryu, S. I. & Shenoy, K. V. Cortical activity in the null space: permitting preparation without movement. Nature Neuroscience 17, 440–448 (2014).
Neuralink devices are investigational and have not been approved by the FDA or other regulatory authorities. This article includes details on the experiences of clinical trial participants which may not reflect the experiences of all participants or future outcomes.