The Unreliable Witness
Individual neurons are unreliable witnesses. Record from the same cell in the mouse visual cortex on Monday, show it a moving image, and it fires vigorously. Return on Thursday with the same clip: silence, or a weak stutter, according to research published in the journal eLife. The single unit does not hold stable information. Yet the mouse sees the same thing both days. The stability is elsewhere.
This is the variability paradox. Individual neurons show variable responses over time, with cells responding vigorously to stimuli on one day but being unresponsive days later, per the eLife study. The human brain contains around 85 billion neurons with some 100 trillion connections between them, as noted in the research. Modern technology can record the activity of millions of neurons simultaneously, the study reports. But researchers can observe only 0.1% of neurons at any given time, according to the findings. The question is whether that fraction is enough.
The Reconstruction
Scientists reconstructed short movies from brain activity of mice that watched video clips, according to work led by Dr. Joel Bauer at the Sainsbury Wellcome Centre at University College London. The method was indirect. Researchers used an infrared laser to record neuron firing in the visual cortex as mice watched 10-second-long movie clips, per the eLife publication. Then they fed blank video data into an artificial intelligence program and altered imagery until it predicted the same brain activity patterns as those seen in mice, as described in the study.
The reconstructed movie clips are grainy and pixellated, the research shows. The reconstructed videos show a pinhole view of the screen the mice see, according to the findings. But they are recognizable. An artificial intelligence program won a recent scientific competition to predict electrical activity in the mouse visual cortex, per the study. The prediction worked despite the chaos at the cellular level, despite recording only a fraction of the neural population, despite the fact that the same neurons would fire differently tomorrow.
The Population Code
Here is the mechanism. Stability does not require reliable units. It requires redundant populations. When one neuron falls silent, others carry the signal. When a cell misfires, the pattern absorbs the noise. The information is not stored in individual firing rates but in the collective activity across thousands of cells, distributed and overlapping.
This is why the reconstruction works from 0.1% sampling. The system is not built on precision at the unit level. It is built on statistical regularity at the population level. The variability that makes individual neurons unreliable becomes irrelevant when averaged across enough cells. The code is in the pattern, not the parts.
Consider what that permits. A brain that tolerates unreliable neurons is a brain that tolerates damage. Cell death from aging, injury, or disease does not immediately degrade perception because the system has redundancy built in. The robustness is not a feature added on top of neural coding. It is the coding itself.
The Sampling Question
We observe 0.1% of neurons and reconstruct what the mouse saw. The videos are grainy, the field of view narrow, but the content is there. Future work could reconstruct an animal's entire field of view using brain activity from both eyes individually, according to the eLife study. The implication is that fidelity does not require completeness.
But what is the remaining 99.9% doing? Three possibilities. First: noise, cellular activity that does not carry information about the stimulus. Second: redundancy, backup copies of the same signal distributed across more cells than strictly necessary. Third: information we are not yet equipped to decode, patterns that matter but that our models cannot yet extract.
The distinction matters. If the unobserved activity is noise or redundancy, then sparse sampling is sufficient. We can scale recording technology without needing to capture every cell. But if the unobserved activity carries information, then our reconstructions are not just grainy. They are incomplete in a way that matters, missing dimensions of perception we do not yet know to look for.
The Human Translation
Several research groups are devising ways to reconstruct images from human brain scans, per the eLife research. A research lab at the University of Bristol is studying how people respond to films using brain activity monitoring, according to Prof. Iain Gilchrist, a neuropsychologist leading the cinema research project at University of Bristol. The cinema lab uses a headset to record brain activity, a heart rate monitor, and infrared cameras to track eye movement and fidgeting, as described by the Bristol team. Researchers measure synchronisation of biometric signals as a sign of audience engagement, per the project.
The mouse work establishes the principle. Perception can be reconstructed from sparse, noisy, variable neural data. The human work is testing the scale. Human brains are larger, more complex, but they operate on the same population coding principles. If the mechanism holds, then what we reconstruct from human brain activity will also be grainy, partial, but recognizable.
The Privacy Boundary
The mouse sees the same image twice. Different neurons fire each time. The reconstruction works anyway. The implication is not that we have cracked the neural code, but that stability does not require the fidelity we assumed. The system tolerates chaos at the unit level because information is distributed, redundant, population-wide.
Which raises the boundary question. If thoughts are stable patterns emerging from unreliable cells, and we can sample 0.1% to reconstruct what you saw, how much do we need to sample to reconstruct what you intended? Perception is passive, a response to external stimuli. Intention is internal, generated without external reference. The mechanisms may differ. But if they do not, if intention is also a population code distributed across redundant, variable neurons, then the same sparse sampling that reconstructs perception could, in principle, reconstruct thought.
The robustness that protects perception from neuronal noise may not protect it from observation. The redundancy that makes the system fault-tolerant also makes it readable from partial data. The question is not whether the technology will scale. The mouse work suggests it will. The question is what we do when sampling 0.1% of someone's brain activity is enough to see not just what they watched, but what they thought about while watching.
