UCL Team Reconstructs Video From Mouse Brain Signals Alone

A team at University College London has done something that reads like a setup for a neuroethics seminar: they took the electrical chatter of neurons in a mouse's visual cortex and worked backward to produce short video clips of what the animal was actually watching.

The work, published in eLife under the title "Movie reconstruction from mouse visual cortex activity," comes from lead author Dr. Joel Bauer at the Sainsbury Wellcome Centre, working with co-authors Troy W. Margrie and Claudia Clopath. The paper has been getting renewed attention this week after a ScienceDaily writeup on September 16 brought it back into broad circulation.

The method starts from single-neuron recordings rather than the whole-brain, region-level signal you get from fMRI. That's a meaningful distinction. <cite index="22-7,22-8">Reconstruction of visual input from human fMRI data has attracted considerable research attention, but comparatively less focus has gone toward vision reconstruction from single-cell recordings, despite its potential to provide a more direct measure of what the brain is actually representing.</cite> Coarser imaging modalities capture population-level activity smeared across millimeters of cortex; single-cell calcium imaging can, in principle, tell you what individual neurons are doing frame by frame.

The UCL team used two-photon calcium imaging to record firing patterns as mice watched video clips, then ran those signals through a dynamic neural encoding model (DNEM) that had won the Sensorium 2023 competition, a benchmarking contest for neural activity prediction. <cite index="12-5,12-6">They reconstructed 10-second video clips based purely on the neural activity of mice, developing an algorithm that translates brain signals back into moving images.</cite> The reconstruction runs at 30 frames per second, matching the original playback rate.

To put a number on accuracy, the authors used pixel-level correlation: comparing each pixel in the reconstructed video against the corresponding pixel in the ground-truth clip. <cite index="12-13">They achieved a pixel-level correlation of 0.57 between ground-truth movies and single-trial reconstructions.</cite> That's not noise, but it's also not a perfect copy. The paper's own reviewers, whose comments are published alongside the article per eLife's open-review model, pushed the team on how to contextualize that figure: one reviewer asked directly whether the encoding model score was impressive and what the theoretical ceiling is given the noise floor of the data.

There are hard limits built into the experimental design. <cite index="13-3,13-4">The imaging data covered a field of view roughly 630 by 630 micrometers in each mouse, covering approximately one-fifth of primary visual cortex V1, so the authors did not expect to get good reconstructions of the entire video frame.</cite> They addressed this by training a mask that highlights the region of the image that the recorded neurons actually care about, based on the collective receptive fields of the recorded population.

<cite index="24-3">The accuracy of the reconstructions improved with the inclusion of data from more individual neurons, the team reported, demonstrating the importance of comprehensive neural data.</cite> That scaling relationship matters: it suggests the approach isn't capped at current performance, but it also means you need increasingly dense recordings to push quality up, which is non-trivial in a live animal.

<cite index="24-4,24-5">To quantify reliability, the team correlated each pixel of the movie between the original and the reconstructed version, finding minimal differences in timing, though they plan to focus on improving resolution and spatial coverage going forward.</cite>

What this isn't is a brain-reading device ready for translation. The model was trained on the same class of stimuli used for reconstruction, which is standard practice in the field but means generalization to arbitrary novel visual scenes remains undemonstrated. The mice also can't report their subjective experience, so the reconstruction shows what neurons encode, not necessarily what perception feels like.

What it is, more precisely, is a proof of concept that single-cell-resolution decoding of continuous visual experience is achievable, at least in a small cortical patch, at least in a rodent. The pipeline from neuron to pixel now closes. Whether that closure scales to other areas, other species, or other sensory modalities is the next question the field will have to answer, probably one replication at a time.

Sources cited:
- eLife, Bauer, Margrie, Clopath (2026) (https://elifesciences.org/articles/105081)
- ScienceDaily, UCL press writeup, September 16 2026 (https://www.sciencedaily.com/releases/2026/09/260914102449.htm)
- Neuroscience News (https://neurosciencenews.com/brain-video-reconstruction-30282/)
- EurekAlert, UCL press release (https://www.eurekalert.org/news-releases/1119258)
- eLife peer reviews (open access) (https://elifesciences.org/reviewed-preprints/105081/reviews)