Trying things out. I will be posting mainly #neuroscience. So here is a question. What is the most important thing we need to be able to do to understand the principles on which brains work?
@dickretired Perturb brains circuits in the way that test state-space concepts of neural activity. E.g. stimulate on particular vectors to, e.g., move along a putative line attractor.
@SussilloDavid @dickretired I second this take from David. I would add, we need this for testing learning algorithms as well, bc we really need to be able to set specific activitystates to see how this induces changes downstream.
@tyrell_turing @SussilloDavid @dickretired Perturbing a system meaningfully requires first knowing the system. In this case a comprehensive #connectome that can serve as the basis for in silico analysis and simulation, and significant enrichment of the molecular properties and how the latter relate to the integration function of each neuron. Otherwise experimentation operates on far too incomplete data supplemented with educated guesses and assumptions.
#neuroscience
@tyrell_turing @albertcardona @SussilloDavid @dickretired good to remember that neural populations don’t know the connectome of their up/downstream partners, and they get along with reading and targeted perturbation just fine.
@dlevenstein @tyrell_turing @albertcardona @SussilloDavid @dickretired Wouldnt regularities in the connectome allow neurons to partially know what their up/downstream partners are connected to?
E.g. through neurons preferentially connecting with neurons with specific connection patterns (https://www.biorxiv.org/content/10.1101/2022.03.09.483605v1.full) or genomic characteristics (https://www.biorxiv.org/content/10.1101/2022.11.06.515380v1)
@achterbrain @dlevenstein @albertcardona @SussilloDavid @dickretired Indeed! I suspect that is actually something used by the brain's learning algorithms for credit assignment.
@achterbrain @tyrell_turing @albertcardona @SussilloDavid @dickretired partially yes! And we can play the same inferential tricks 😁
@albertcardona @tyrell_turing @SussilloDavid @dickretired Do you think this level of detail is as necessary in larger more complex brains? It seems if state spaces offer the right level of explanation then individual nodes of the network could rightfully be abstracted away.
@beneuroscience @tyrell_turing @SussilloDavid @dickretired Seems premature to dismiss data as unnecessary only because it’s hard to acquire.
@albertcardona @tyrell_turing @SussilloDavid @dickretired Not dismiss and maybe not unnecessary, but perhaps less of a priority? Just curious whether you think the level of explanation/abstraction should scale with the network complexity.
@beneuroscience @albertcardona @tyrell_turing @SussilloDavid
Good question. It can’t. What we need to understand are the principles. We will never be able to comprehend a working human brain. Drosophila, Angel Fish perhaps.
@dickretired @albertcardona @tyrell_turing @SussilloDavid If principles == shared principles among all brains, OK. Alternatively, new principles may emerge in more complex brains, and revealing them may require more abstraction.
Agreed. But evolution tends to use what has already worked. so studying nematode worms, drosophila and angel fish (as some are) may be a good place to start.
@beneuroscience @albertcardona @tyrell_turing @dickretired
Sidestepping the connectome question, it's worth noting that if perturbation vectors are directly aligned to neurons, that's *really useful.* It's even more useful if they are aligned to neurons of a specific cell-type. So state space is useful, but the details really matter, too, IMO.
E.g. see https://www.sciencedirect.com/science/article/pii/S0092867422011138
@beneuroscience @albertcardona @tyrell_turing @dickretired
In this paper, we found an optotagged cell class in the medial habenula, which by all appearances looks like a line attractor. If we could stimulate the population of neurons (and some math simplifies), then we'd be one step closer to understanding integration in (mouse) brains.
@SussilloDavid
Sounds very cool and it's been on my reading list, thanks :)
@albertcardona @SussilloDavid @dickretired I partially agree. Without the connectome you will need to make some assumptions, but I think the advancements in understanding that could be possible with fine-grained perturbations sans connectome would be still quite large.