UCL Reconstructs Videos From Mouse Brain Activity — SkimNews
Get the Health newsletter
Daily health & science — research, biotech, public health, the studies worth knowing. Free.
- UCL researchers led by Dr. Joel Bauer at the Sainsbury Wellcome Centre reconstructed 10-second videos from mouse visual cortex activity, producing clips that resembled scenes the model had never been trained on.
- The team used single-neuron calcium imaging rather than broader fMRI signals, recording activity from individual brain cells to capture a more detailed picture of visual representation.
- The method adapted a dynamic neural encoding model from the 2023 Sensorium Competition, subtracting predicted "blank screen" responses from actual movie-watching activity so an algorithm could iteratively adjust pixels of a blank video.
- The eLife study (DOI: 10.7554/eLife.105081.3) found reconstruction accuracy improved as more individual neurons were included, indicating the model was inferring visual information rather than memorizing prior clips.
- Researchers evaluated results via pixel correlation, finding only small timing differences between originals and reconstructions but acknowledging that image resolution and scene coverage still need improvement.
- Dr. Bauer said the brain's internal representation "skews and warps" sensory information — not as an error but as a feature of perception the team now wants to map using this technique.
Why it matters: The technique moves visual decoding from coarse fMRI signals to single-cell resolution in mice, giving researchers a direct window into the gap between the world in front of an animal and what its brain actually represents. With accuracy scaling alongside neuron count, denser recordings could let scientists pinpoint which visual features the brain emphasizes, filters, or distorts.
Ask SkimNews
