year: 2024/08
paper: https://arxiv.org/pdf/2408.10234
website:
code:
connections: consciousness,
- conscious processing speed is at ~10bit/s pretty much across many experiments.
- We receive bits more information than we cosnciously process.
- Photographic memory is not really a thing.
- “Can’t rlly test unconscious processing (sth sth)” (i kinda disagree with this and think they’re giving the unconscious processing too little credit).
- ~50TB required to specify all synapses (nurture)
- <1GB for the genome (nature)
- Authors think BCIs will fail or like won’t be able to extend our communication throughput past 10b/s. I disagree. I don’t think the conscious processing we measure(d) so far is the only thing that matters…
- “Why is conscious cognition not sequential and not parallel like much of the rest of our processing?” (→evolutionary histroy maybe)
neurons are not really noisy + do actually transmit quite a bit of information + there is not really much redundancy in brain structure
Another argument one hears frequently is that neurons are simply inefficient devices for information processing, so nature needs to employ huge numbers of them to complete even simple operations. In particular, individual neurons – unlike transistors – are subject to random biochemical noise, so perhaps one needs to average over many neurons to get a reliable signal. Also, perhaps the brain includes a large amount of redundancy, with populations of essentially identical neurons, to guard against the loss of neurons from aging or injury. Neither of these explanations seems plausible.
Regarding “noisy neurons”, the information-theoretic analysis reviewed above (section 3) gets at exactly that point and shows that a single neuron can reliably signal several bits per spike about its dendritic inputs. As a rule, once experimenters discover what a particular neuron encodes, they realize that the representation is amazingly precise. A single optic nerve fiber can reliably signal the arrival of its preferred feature in the visual stimulus with a precision of 1 millisecond [82]. Even deep in the primate cortex, a single neuron can distinguish between two visual stimuli as precisely as the entire monkey [83]. There is increasing recognition in the community that what we used to call noise in neural recordings is actually a reliable signal about variables that the experimenter doesn’t control [84, 85]. On the subject of redundancy, again there is precious little evidence for duplication of essentially identical neurons. For example, in the fruit fly, many neuron types exist in just two copies, one on each side of the brain [86, 87]. One might argue that fruit flies are cheap and short-lived, so nature simply duplicates the entire organism to ensure robustness. However, the principle applies even to large animals like ourselves. In the primate retina, every point in the visual field is covered by just one neuron of each cell type, with very little overlap [31, 88, 89]. In the visual cortex, a small stroke can abolish vision in a section of the visual field. Apparently, there is no redundancy built in for one of the most common forms of brain injury.