Learning to decode neural activity without every label

What the work claims

Neural decoders for brain-computer interfaces and closed-loop experiments have traditionally relied on supervised learning, which means every training example needs a paired behavioural label. That requirement is expensive and limiting. Mao and colleagues propose MOJO, short for Masked autOencoder-based JOint training, which combines a supervised decoding objective with a self-supervised objective that predicts masked spike counts from latent representations.¹

The claim is that this joint training improves decoding across species, brain regions, tasks, and recording modalities, with the largest benefits when labels are scarce. The paper tests this on three spike datasets: non-human primate reaching, mouse visual and decision-making recordings, and human electrocorticography during speech articulation. MOJO is evaluated on top of two existing spike-tokenizing backbones, POYO and POSSM, and adds only a small number of parameters: roughly 10 percent for the POSSM backbone and 2 percent for POYO.

How it works

MOJO starts from the POYO family idea that each neural spike can be treated as a token carrying the identity of the recording unit and the exact time of the spike. A cross-attention encoder turns sparse spike sequences into latent tokens. The original POYO models then run a supervised decoder on those latents. MOJO adds a second decoder head that tries to reconstruct spike rates for time intervals that were masked during encoding.¹

The masking is applied in latent space, not at the raw spike input. During training, a Bernoulli mask with probability 0.5 replaces all latent tokens in selected time chunks with a learned mask vector. The self-supervised head then predicts the Poisson spike rate for each unit in the masked chunks. The total loss is a simple sum of the supervised decoding loss and the spike-reconstruction loss, with both coefficients set to one. Because the backbone parameters are shared between the two pathways, the extra parameter cost is small.

The supervised task varies by dataset. For monkey reaching the target is two-dimensional hand velocity. For mouse vision it is classification of natural scenes or drifting-grating parameters. For mouse decision tasks it is choice, block context, wheel velocity, and whisker motion. For human speech it is classification of consonants, vowels, and full syllable pairs. The architecture is the same across all of them except for the final task-specific readout heads.

Where a skeptic should push

The strongest result is also the most carefully bounded. MOJO's unit-identification finetuning, which updates only session and unit embeddings, can match the fully supervised finetuning of a purely supervised POSSM model while training fewer than 18,000 parameters instead of more than 7.6 million. That is a real efficiency gain, but it applies to the readout, not to the brain. The tissue itself is not learning; the model in silicon is.

The few-shot numbers are promising but need context. With two labelled trials and up to 32 unlabelled trials, MOJO reaches more than 60 percent of fully supervised performance; with four labelled trials it reaches more than 75 percent.¹ But "fully supervised performance" here is on the same small calibration set, not on a clinically relevant generalisation test. The model is still subject-dependent and session-dependent; it learns to read a particular electrode configuration on a particular day.

The interpretability results are suggestive but not causal. MOJO's unit embeddings can predict brain region and spike statistics better than purely supervised embeddings, and embedding similarity correlates with electrode distance. That tells us the representations are structured, not that the structure is necessary for decoding performance.

Finally, the authors note that MOJO is data-hungry. If the unlabelled dataset is too small, the self-supervised objective can hurt rather than help. And when finetuning to a new session, unit embeddings are relearned from scratch, which wastes the pretraining investment. For organoids, where every culture is a new non-transferable object, that limitation is acute.

Self-supervised decoding as an organoid readout

Every organoid computing experiment has the same two-part architecture: living tissue generates activity, and a conventional computer reads and interprets it. MOJO addresses the second part. It shows that the readout model can be pretrained on large volumes of unlabelled neural activity, then adapted to a specific task with only a handful of labelled examples. That is a direct playbook for organoid intelligence: record activity from many cultures and many idle conditions, pretrain a spike-tokenizing foundation model, and finetune it only when you need the culture to produce a particular output.

The non-obvious implication is about where the learning lives. MOJO's gains come entirely from the silicon side. The neurons are not trained; the decoder is. This is consistent with nearly every current organoid computing demo, in which the culture provides a rich dynamical substrate and the experiment designer supplies the task, the feedback, and the learning algorithm. MOJO makes that division of labour more data-efficient, but it does not dissolve it. If the long-term goal is for organoids themselves to learn and retain computation, a decoder that adapts in silicon is a useful crutch, not a solution.

The opportunity is scale. Organoid recordings are currently fragmented across labs, protocols, and developmental time points. A MOJO-style approach could in principle pool unlabelled recordings across experiments to pretrain a general organoid activity model, then finetune on the specific culture and task. The spike-tokenizing input format is flexible enough to handle different electrode counts and configurations without redesigning the network. The caveat is biological variability: organoids differ far more than sessions within one species, so cross-culture transfer is a harder problem than the cross-session transfer shown here.

The threat is expectations. MOJO's results are strong enough that they may be read as evidence that decoding is becoming a solved problem. It is not. The paper works with spike-sorted single units from invasive recordings, clean trial structure, and curated tasks. Organoid readout will face unsorted multi-unit activity, slower dynamics, less stereotyped responses, and no reliable behavioural labels at all. MOJO points a direction, but the path from here to an organoid readout model is longer than the paper's headlines suggest.

The bottom line

Established: joint self-supervised and supervised training of spike-tokenizing models improves neural decoding across multiple species, tasks, and modalities, with the largest gains when labels are limited. Established: the method adds only a small parameter overhead and produces more interpretable unit embeddings. Not established: that the same transfer works across the much larger biological variability of organoid cultures, or that the method removes the need for silicon-side learning.

For organoid intelligence, MOJO is a useful readout roadmap. It suggests that the field should invest in pooled, unlabelled organoid recordings and foundation-style decoder pretraining, rather than treating each culture as an isolated supervised-learning problem. The deeper question it leaves open is whether the organoid itself can ever become the learner, or whether biological computing will remain a substrate that silicon learns to read.