Science 1 publisher2 min readPublished September 16, 2026 at 8:42 AM PDT

UCL team rebuilds 10-second movies from single-neuron recordings in mouse visual cortex

Because the recordings come from identified single cells, the number of neurons feeding a reconstruction is something the experimenter controls, and the output can be laid beside the movie the mouse actually watched.

The Scientist · Science desk

Photograph accompanying UCL team rebuilds 10-second movies from single-neuron recordings in mouse visual cortex — Photo: ucl.ac.uk

What happened

Researchers led by University College London rebuilt 10-second video clips using only the recorded activity of neurons in the visual cortex of mice.
The input came from individual brain cells rather than the broader brain imaging signals that human decoding studies have used with fMRI.
The encoding model at the centre of the method was built by a different research team for the 2023 Sensorium Competition to predict how single neurons respond while mice watch movies.
After training, the team recorded a mouse watching a video that had been kept out of the training data and reconstructed a 10-second clip resembling it from that activity alone.
The activity itself was read with a microscopic imaging method that marks an active cell by the localized rise of calcium inside it.

Compiled by The ScientistSomething wrong?How this is made

Why it matters

capability With the input measured cell by cell, how many neurons feed a reconstruction becomes a variable the experimenter sets, and Bauer reports accuracy climbing as more cells are added.
constraint Every reconstruction passes through the encoding model, so the model's blind spots would be read as the brain's unless a lab checks them separately.
precedent The stated ambition of comparing how different species see the same surroundings now depends on collecting comparable single-neuron recordings in each of them.

Decoding here runs an encoder backwards. The team first asked the model how the recorded neurons would behave in front of a blank screen, then compared that prediction with what those neurons actually did while the mouse watched a movie ^[7]. The difference drove an algorithm that edited the pixels of an initially blank movie, and with each adjustment the reconstruction became more like the clip the animal had been shown ^[9]. So the output is the movie that best explains the recording under that particular model of those neurons. A visual feature the encoding model cannot express will not appear in the reconstruction, whatever the cortex did with it.

The model's origin matters for the same reason. It came out of the 2023 Sensorium Competition, and its job was predicting how single neurons respond as a mouse watches a movie, with the animal's movements and pupil diameter among the inputs ^[6]. The UCL group refined the approach on the same dataset ^[7].

Joel Bauer, the lead author, is at the Sainsbury Wellcome Centre at UCL ^[16]. He said the field needed a more general tool: "The current methods of understanding what specific groups of neurons are representing are not very generalizable to situations which haven't been specifically tested for" ^[12]. His group's interest is the gap between what sits in front of an animal and what the cortex holds, and mapping that gap could show which visual features get emphasised, altered or dropped ^[14]. The work is published in eLife ^[2].

On the held-out clip, Bauer said: "The accuracy of the reconstructions improved with the inclusion of data from more individual neurons, demonstrating the importance of comprehensive neural data" ^[13]. The reported improvement comes without a stated neuron count. The held-out test shows that the recorded activity carried enough information about a scene the model had never been trained on for the pipeline to infer it, so the system was not returning a memorised video ^[11].

The transformation from scene to internal representation is the target this method was built to attack, and the study demonstrates the method ^[1]^[14]. A reconstruction of this kind is a statement about the information in the cells that were recorded, not about what the mouse experienced.

What to watch

Whether reconstruction quality keeps rising with neuron count or saturates, and at what population size.

Claim ledger

Ranked by verification strength, evidence, and original report placement.

[13]
Dr Joel Bauer said: "Using this approach, we were able to achieve high-quality reconstructions of 10-second video clips. The accuracy of the reconstructions improved with the inclusion of data from more individual neurons, demonstrating the importance of comprehensive neural data."
ReportedSupportedSource: Dr Joel Bauer, Sainsbury Wellcome Centre, UCL2 sources— sign in to open View cited source
[1]
Researchers led by University College London reconstructed videos from the brain activity of mice, producing 10-second clips based only on signals from neurons in the visual cortex.
ReportedSupportedView cited source
[2]
The findings are published in eLife.
ReportedSupportedView cited source
[3]
The ScienceDaily release describing the work is dated September 16, 2026 and credits University College London as the source.
ReportedSupportedView cited source
[4]
In human studies, scientists have shown people images and movies while recording brain activity with fMRI, then used those measurements to try to reconstruct visual information down to the level of individual pixels.
ReportedSupportedView cited source
[5]
Instead of relying on broader brain imaging signals, the researchers used recordings from individual brain cells in mice; these single-cell measurements can provide a more detailed picture of how visual information is represented in the brain.
ReportedSupportedView cited source
[6]
To reconstruct the movies the team used a dynamic neural encoding model originally developed by another research team for the 2023 Sensorium Competition; the model is designed to predict how individual neurons will respond while mice watch movies and also takes into account other factors including the animals' movements and changes in pupil diameter.
ReportedSupportedView cited source
[7]
The UCL researchers refined the approach using the same dataset: they first calculated how the neurons were predicted to behave if the mouse had been looking at a blank screen, and compared that prediction with the neurons' actual activity while the mouse watched a movie.
ReportedSupportedView cited source
[8]
The neural activity had been measured with a microscopic imaging method that identifies which individual brain cells are active by detecting localized increases in calcium levels.
ReportedSupportedView cited source
[9]
Using the difference between predicted and measured activity, an algorithm gradually changed the pixels of an initially blank movie, and with each adjustment the reconstructed video became more similar to the one that had actually been shown to the mouse.
ReportedSupportedView cited source
[10]
After the model had been trained, the researchers recorded a mouse's brain activity while it watched a video that had never been included in the model's training data, and using only that neural activity the system reconstructed a 10-second movie resembling the unseen video.
ReportedSupportedView cited source
[11]
The result suggests that the method was not simply memorizing previously shown videos, but was able to use patterns of neural activity to infer visual information from a new scene.
ReportedSupportedView cited source
[12]
Lead author Dr Joel Bauer said: "We wanted to have a better way of investigating how the brain interprets what we see. The current methods of understanding what specific groups of neurons are representing are not very generalizable to situations which haven't been specifically tested for. And so, we wanted to develop a method that can capture what is being represented in the brain and compare that to reality."
ReportedSupportedSource: Dr Joel Bauer, Sainsbury Wellcome Centre, UCLView cited source
[14]
The researchers are especially interested in the differences between what is physically present in front of an animal and how that information is represented inside the brain; studying those differences could reveal which visual features the brain emphasizes, changes, or filters out.
ReportedSupportedView cited source
[16]
Dr Joel Bauer, the lead author, is at the Sainsbury Wellcome Centre at UCL.
ReportedSupportedView cited source
[15]
The findings may eventually make it possible to compare how different species perceive the same surroundings.
ReportedInsufficientView cited source

Sources

Publishers with included, body-backed reporting in this cluster.

sciencedaily.com
1 article · September 16, 2026
Scientists recreated what mice saw from brain activity alone

Ask Clarity

Have a question this piece left open? The answer comes from its reporting and claim ledger — nothing else.

Try

Clarity analysis

How the coverage leans and what the record actually supports. Read the methodology

Perspective Coverage

How much of the coverage speaks to builders, operators and investors. It measures emphasis, not whether the story is right.

UCL's method is worth walking through in detail. UCL took a forward model from the 2023 Sensorium Competition, predicted how each imaged neuron would behave in front of a blank screen, compared that with the calcium-imaging traces recorded while the mouse watched a film, and used the mismatch to push pixels around until the predicted activity matched the measured activity. The key test was reconstructing a held-out clip: a video the model had never trained on, reconstructed for ten seconds from neural activity alone. Bauer said accuracy "improved with the inclusion of data from more individual neurons". ScienceDaily's copy of the release does not give neuron counts or pixel-correlation scores, so the eLife paper is the only place to check that scaling claim. Because the pipeline runs on a public competition dataset, it is reproducible outside the lab that ran it.

Reality

Evidence, adoption, hype gap, incentives and confidence, each read independently from the current record on the scale shown beneath it. How these are measured.

Evidence48

0100

One institutional account, checkable only in eLife

The method is described in enough detail to be recognisable: a Sensorium 2023 forward model, a blank-screen baseline, then iterative pixel updates driven by the gap between predicted and measured calcium signals. Reconstructing a clip the model never trained on is the right test to run, and the release reports it. Neuron counts and pixel-correlation scores are missing from the release, so the quality and scaling claims rest on Bauer's own sentences, and the eLife peer review is the place those claims can actually be checked.

AdoptionInsufficient

Nothing to count yet

This is one lab's result on one shared dataset. No other group is reported using the reconstruction method, and there is no tool, code release or product for anyone to pick up. The reuse in the story runs the other way: UCL adopting a model and recordings from the 2023 Sensorium Competition.

Hype gap+15

Topics and entities

Follow any of these and your For You feed starts watching them — no settings page required.

Topics

Entities

science
Janelia's WHOLISTIC reads calcium signals from nearly every cell in a live zebrafish larva
1 publisher
science
Chemoradiotherapy reshapes which mutant clones win in the normal esophagus
1 publisher
science
Only gentle terrain fits beneath Chile's 21.9-million-year-old ash sheet
3 publishers
build
Clinicians ask for chatbot screening before psychiatry rules on AI psychosis
1 publisher

What happened

Why it matters

What to watch

Claim ledger

Sources

Let Clarity write the post for you.

Topics and entities

Topics

Entities

Related stories

Claim ledger

Let Clarity write the post for you.

Topics and entities

Topics

Entities

Perspective Coverage

Reality

Related stories

Sources