I’d love to hear your thoughts on this. I recently watched a talk by Dmitri Chklovskii, where he discussed how his team mapped the entire neural circuitry of a worm and attempted to simulate its behavior—only to find that the results fell short. He suggests that neurons might be more sophisticated than we currently assume, proposing that a single neuron could function as a feedback controller. I also came across a talk by Gabriele Scheler, who shares a similar view that neurons are smarter than existing models suggest. She’s working on a model that incorporates intracellular processes and explores how neurons might self-program. You can find their papers here:
• Chklovskii et al.: https://www.biorxiv.org/content/10.1101/2024.01.02.573843v1
• Scheler: https://arxiv.org/pdf/2209.06865
Do you think these ideas are significant for designing something like a neocortical circuit? I’m curious about your opinions!
Hey @James, just sharing my two cents here too. Interesting question!
I have some interest in connectomics and I think I heard a phrase along the lines of “a dirty secret in computational neuroscience is that we had the connectome of C. elegans since mid 1980s, yet still cannot simulate a complete C. elegans”. (This isn’t to discourage mapping efforts for more complex connectomes like mouse or human brains—quite the opposite.)
I think the papers you shared are indeed valuable for designing neocortical circuit models!
I had this thought experiment:
Let’s say we had:
A full connectome (complete wiring + synapses + receptor types, etc.)
Full biophysical properties (ion channel densities, pathways, etc.)
Enough computational power to stimulate every neuron with Hodgkin-Huxley level detail, including molecular dynamics.
While this may theoretically enable full-resolution brain simulation, I believe this primarily represents “copying” rather than “understanding” the brain (though this would still be an extraordinary achievement).
I think models proposed by Chklovskii and Scheler would still be valuable even in this scenario because I think these tools/models would help us better understand, interpret, design, and/or engineer neural networks with desired behavior.
Just my two cents. Thanks for sharing those papers.
Interesting question!