October 19, 2026
Technology Watch

From the laboratory to industry: following new technology on its way

← Front page

arXiv

Neural representations outlast lost neurons

Image drawn by AI

NeuroLens could make implanted speech and motor interfaces less dependent on changing neural recordings, while leaving a harder scientific question unresolved. Its authors trained the model on 53 daily mouse sessions and tested whether it preserved task-relevant representations across time. In a rehabilitation clinic, the result would be a short calibration exercise before attempted speech becomes text, provided the representation holds across many implants, users, tasks, and longer periods.

Every day, a brain–computer interface has to face a small version of the same problem: the brain has not necessarily changed, but the recording has. Electrodes can drift, their signals can weaken, and the nearby neurons they pick up can change. A system that turned electrical activity into cursor movements or text yesterday may therefore need new calibration today. For scientists, the problem is harder still. A different signal might mean the animal is learning—or simply that the electrode is now listening to different cells.

The unstable window into the brain

An implanted electrode does not record the brain as a whole. It records activity from a changing group of neurons. Yet the same behavior can continue even when that group changes, because many neurons can help produce the same result.

Most neural models try to reconstruct or predict the recorded activity itself. That can preserve useful information, but it can also preserve the noise and quirks of that day’s recording. A model tested on later days faces a tougher test than one tested on new examples from the same day: the input has shifted.

This matters especially for speech interfaces. Decoding a phoneme—a small sound unit such as part of a spoken word—is not the same as decoding the meaning of a whole sentence. A useful speech system must capture what a person intends to say, not only short-lived movement patterns in the brain.

A model that keeps the meaning

NeuroLens takes a different approach. Instead of trying to reproduce every recorded signal, it predicts what the brain’s future internal representation will be. The representation is a compact pattern that aims to keep information about the task while dropping details tied to a particular recording day.

The model first uses short calibration trials to build a signature for each recorded neuron. That signature describes how often the neuron fires, the timing pattern of its firing, and how closely it works with the rest of the population. A day-adaptive encoder then uses those signatures to place changing groups of neurons into a shared space. In that space, the model can compare activity from different days without treating every changed neuron as a new computation.

The authors trained NeuroLens without requiring labels for every example. Its prediction system looks at past latent states and predicts future ones. An added constraint is meant to stop all inputs from being reduced to the same uninformative constant.

In tests using 53 daily sessions from one mouse over roughly three months, the model was trained on earlier days and tested on later ones. The authors report better decoding of the animal’s left-or-right choices than a comparison model, with a larger gain for choice than for movement. In recordings from two people with ALS-induced anarthria who attempted to speak displayed sentences, NeuroLens improved decoding of sentence meaning more than decoding of phonemes. The reported gain for sentence-embedding decoding was 38%, compared with 4% to 16% for phoneme decoding.

Refer to caption
Figure 1: NeuroLens framework. (A) Model overview. A day-adaptive encoder maps spike counts at each time bin to latent representations, and an autoregressive (AR) predictor shared across days predicts future latents from past latents. The model is trained using a latent prediction loss together with SIGReg loss to prevent representation collapse. (B-C) Day-adaptive cross-attention encoder based on calibration data. At each time bin, an MLP independently projects each neuron’s spike count into a neuron token. These tokens serve as the keys and values of the cross-attention module. Because attention is permutation-invariant, we distinguish neuron identities using calibration trials sampled from each recording day. From these trials, we compute firing-rate statistics, autocorrelograms (ACGs), and population coupling to generate a signature for each neuron. A second MLP maps these signatures to neuron embeddings, which are concatenated with brain-region embeddings and added to the corresponding neuron tokens. Learned queries then aggregate the variable-length sequence of neuron tokens through cross-attention, producing a fixed number of latent embeddings for the AR predictor. Processing each time bin independently prevents temporal leakage, while calibration-based neuron embeddings enable adaptation to previously unseen neural populations across recording days.## 4 Methods Source: NeuroLens: Learning Latent Embeddings of Neural Semantics from Chronic Recordings（arXiv）

What a steadier interface could change

In a rehabilitation clinic, a person with an implanted speech interface could begin a session with a short calibration exercise instead of a lengthy retraining routine. If the shared representation remains reliable, the system could continue turning attempted speech into text even as the recorded neurons change. The same idea could reduce recalibration for implanted motor interfaces, such as systems that control a cursor.

For neuroscientists, the benefit would be different. They could compare population activity across weeks or months without assuming that every lost or shifted neuron proves the brain’s underlying computation has changed. That could make long-running studies of learning easier to interpret.

But NeuroLens does not yet settle the most important scientific question. The method cannot separate changes caused by learning and behavior from changes intrinsic to the brain. In the human cross-time tests, few-shot adaptation performed poorly, so the models had to be fully fine-tuned using each participant’s training data. The authors’ proposed future—a short calibration rather than lengthy retraining—depends on the representation working across many implants, people, tasks, and longer periods than those tested here.

Refer to caption
Figure 3: NeuroLens learns neural representations that generalize across time, enabling rapid adaptation to unseen neural populations and more stable decoding than SOTA baselines. (A) The mouse makes increasingly accurate decisions about the visual stimulus location from April to June, while the number of simultaneously recorded neurons decreases over time. We pretrain on the first 39 days and test generalization on the held-out last 14 days. (B) Across 14 held-out days, NeuroLens outperforms baselines in choice and movement decoding, demonstrating better generalization across time. (C) Comparison of predicted movement-related behavior averaged across all trials (top) and predicted behavior from single trials (bottom), shown to demonstrate that the baselines can predict trial-to-trial variability in behaviors. The R2R^{2} values for both trial-average and single-trial behavior are shown. The dashed line shows the stimulus onset. (D) T12 and T15 show clear day-to-day distribution shifts in neural activity. We hold out the last 5 and 9 days, respectively, for testing and pretrain on the remaining days. (E) NeuroLens (FT, full fine-tuning) consistently outperforms baselines in phoneme and sentence embedding decoding across 5 (T12) and 9 (T15) held-out days, reflecting better generalization over time. Lower phoneme error rate (PER) and higher top-5 sentence retrieval accuracy indicate better decoding performance. (F) Latent trajectories for different phoneme onsets, projected onto the first three PCs, across held-out days in T15. Dark (light) lines show averaged (individual) time windows. NeuroLens shows clearer phoneme separation than causal NDT, indicating better held-out adaptation, quantified by LDA balanced classification accuracy.### 5.2 Generalization Across Time Source: NeuroLens: Learning Latent Embeddings of Neural Semantics from Chronic Recordings（arXiv）

Sources

arXiv, “NeuroLens: Learning Latent Embeddings of Neural Semantics from Chronic Recordings”