Post

Dan Goodman
@neural-reckoning.org

How can we test theories in neuroscience? Take a variable predicted to be important by the theory. It could fail to be observed because it's represented in some nonlinear, even distributed way. Or it could be observed but not be causal because the network is a reservoir. How can we deal with this?

8 reposts58 likes

Alfred Nobel
@alfrednobel.bsky.social

Imo you will never be capable to solve anything on the neuro-side without an explanatory model irt all functional/behavioral perception processes or the beginning of such a model. And even then there is "a long and bumpy road in front of you":

Dan Goodman
@neural-reckoning.org

Increasingly feel like this isn't a theoretical problem but a very practical one that comes up all the time. I'd be interested if anyone has seen anything practical that addresses this.

Michael W. Cole
@mwcole.bsky.social

Once you have a flow model that generates a phenomenon of interest, you can lesion each empirical constraint (e.g., each connection or task-evoked activation) to determine which contributed to generation of that phenomenon. Follow-up empirical stimulation or lesion work can further verify this.

Blake Richards
@tyrellturing.bsky.social

I think it requires computational neuroscientists to bite the bullet and commit to specific hypotheses about how certain variables are handled physiologically. Note: though I try to do this sometimes, this critique applies to a good deal of my own work!

Dan Goodman
@neural-reckoning.org

Yes to comp neuro biting the bullet, but I'm still not entirely sure if this is the bullet we need to bite. I certainly wouldn't want to say everything has to be done with multicompartmental HH neurons or anything like that.

Blake Richards
@tyrellturing.bsky.social

No, but that's not what I mean. Rather, I mean things like actually saying "Our model predicts that this cell type computes X" or "this neuromodulator signals Y". Example: I think the Schultz, Dayan, Montague work was great for laying bets on a specific role for the VTA dopamine neurons.

Andrew MacAskill
@macaskillaf.bsky.social

Similar to posts before - doing this in a 'many lines of evidence' approach too is probably ideal - no one golden expt. Just a series of thought out experiments designed to test the limits of a theory, each adding weight to the next. saying this out loud makes me feel old...

Dan Goodman
@neural-reckoning.org

It probably did and also separate issues. But, focusing on behaviour without worrying about how that behaviour is implemented at the low level maybe avoids the problem to some extent?

Nicholas Guttenberg
@ngutten.bsky.social

It suggests to me that maybe the natural invariants exist at a higher level of abstraction, so if you operate at the correct level of abstraction you'd hopefully find that the ambiguous possibilities all coarse-grain to the same higher-level variable.

Nicholas Guttenberg
@ngutten.bsky.social

Sort of like how you might be able to talk about e.g. whether a sorting algorithm is O(N) or O(N^2) independent of many, but critically not independent *all* of the details of how it's actually implemented. And sometimes you can prove things are possible/impossible independent of implementation.

Etienne Thoret
@etiennetho.bsky.social

To avoid the problem to some extent, should neuroscience focus on the neural system without worrying about how that neural system is implemented at the low level? This is what the Tortoise said to Achilles 🙃 & I don't know what this would mean either for neuroscience: not to focus on implementation?

Yohan J John
@dryohanjohn.bsky.social

A historical lens might be useful here. How was the neuron doctrine tested? How was the chemical vs electrical synapse debate settled? Looking outside neuroscience helps too. How was the modern synthesis tested? Also: are you distinguishing between theories and models?

Yohan J John
@dryohanjohn.bsky.social

Pretty sure the neuron doctrine has won out over Golgi's reticularism. 😅 And the chemical neurotransmitter debate is quite interesting ("soup vs sparks"), because one side used theoretical arguments that chemical transmission would be too slow. anatomypubs.onlinelibrary.wiley.com/doi/10.1002/...

anatomypubs.onlinelibrary.wiley.comThe History of the SynapseWhy did I choose this particular topic for my lecture rather than the history of neuroscience or the history of the neuron? Simply because I believe that every disciple has the obligation to pay homa....
Dan Goodman
@neural-reckoning.org

Well there are clearly cells but there are also more soup like mechanisms at play (neuromodulators). I think it's still an open question how important these components are and what their precise functions are, no?

Shahab Bakhtiari
@shahabbakht.bsky.social

This is probably a great example of what you want (a finding a theoretical variable in the brain) that panned out through >two decades of research. I don't think it's infeasible in principle, but it just takes time and a lot of back-and-forth. bsky.app/profile/tyre...

Blake Richards@tyrellturing.bsky.social

1/3) This may be a very important paper, it suggests that there are no prediction error encoding neurons in sensory areas of cortex: www.biorxiv.org/content/10.1... I personally am a big fan of the idea that cortical regions (allo and neo) are doing sequence prediction. But... 🧠📈 🧪

biorxiv.orgSensory responses of visual cortical neurons are not prediction errorsPredictive coding is theorized to be a ubiquitous cortical process to explain sensory responses. It asserts that the brain continuously predicts sensory information and imposes those predictions on lo...
Dan Goodman
@neural-reckoning.org

Without reading the linked paper I don't know, but if it's saying they recorded a bunch of neurons and didn't find X that's exactly the sort of test that I was talking about. It's hard to be convinced because we know the brain doesn't need to represent info in a way that's transparent to us.

Shahab Bakhtiari
@shahabbakht.bsky.social

Right, but PC has become one of the best algorithmic theories that has moved gradually closer to clear hypotheses about its neural implementations (e.g., error coding by sensory neurons). This paper seems to reject at least one such hypothesis.

Alessandro Galloni
@argalloni.bsky.social

The closest I can think of is something like the efficient coding hypothesis, where we can make a theory about representation that would be optimal for encoding sensory info in a given environment, and predict how this should change when sensory statistics are altered

Dan Goodman
@neural-reckoning.org

Representation is always tricky though, since we normally need to do linear readout to find them, and linear readout can do almost anything if you have an embedded reservoir (quite likely in my opinion).

Richard Lange
@bonsai-lab.bsky.social

We can falsify hypotheses about particular vars being represented by particular neurons in particular formats. Marginalizing a theory prediction over all possible formats/implementations sounds, uh... like a good problem for "future work" 😬

Tim Verstynen
@tdverstynen.bsky.social

Deborah Mayo’s work touches on this. Here is the one that is the closest to addressing your question. I haven’t read her newer work (on my list), but they may as well. press.uchicago.edu/ucp/books/bo...

press.uchicago.eduError and the Growth of Experimental KnowledgeWe may learn from our mistakes, but Deborah Mayo argues that, where experimental knowledge is concerned, we haven’t begun to learn enough. Error and the Growth of Experimental Knowledge launches a vig...