(And no I'm not trying to reopen any old #manifoldsplaining nonsense)
Show this thread
Question for neuro people: can you say concretely what you feel this style of analysis provides? I too think it has been important for my thinking, but I can't exactly say why that is or what really crucial advance it has offeredhttps://twitter.com/elibpollock/status/1080944198369579014 …
-
-
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
To me it marks a shift away from thinking about the response properties of single neurons as the measurement of interest, instead viewing each neuron as a noisy observation of the population state.
-
Fully agree with Eric' & Chetan's insightful comments. Just add that this view helps us move away from the IMO misleading search of lawful single neuron representations: I don't think there are kinematic neurons in M1 or force neurons in S1: those arise from sporadic correlations
-
Information or computations are population-wide and single neuron representations probably *clears throat* just epiphenomenal. But of course Fetz already said it in... 1992!!!
-
Yea its important to de-throne single neurons & focus on a population view. But why does kinematic encoding have to go with it? It seems like taking the population view means giving up on describing the computation in concrete terms. Its just saying the dynamics "match" behavior
-
Or at least that's what I worry about it. Of course there's no reason neurons have to be using the same principles as engineers, so maybe it was wrong to try to describe the computation in those terms anyways. But I want to try to describe it somehow!
-
Taking the population point of view in fact enables you to describe computations over distributed representations. Here is a framework for thinking of computations in that way:https://www.cell.com/neuron/pdf/S0896-6273(18)30543-9.pdf …
-
Love this paper! One of my favourite reads of 2018...
End of conversation
New conversation -
-
-
Really interesting (and important) question. Re: "Low-D/state space is good for visualization" - sure, certainly useful as a tool to then ask other scientific questions. But there are also key scientific insights from low-D/state-space approaches. 3 motor systems examples:
-
1st,
@MattAntimatt , 2014 NN (Cortical activity in the null space, with@shenoystanford). Showed that activity during movt preparation lives in an orthogonal subspace from activity that correlates with muscle activation.pic.twitter.com/DndFgKccyc
-
Not just visualization, and certainly not a given that the system must be organized this way (could've been distinct cell types, or gating, etc). Rather, suggests low-D / state space is important to how the network implements computations.
@gamaleldinfe had a great follow-up. -
2nd example: Sadtler et al., 2014 Nature (Neural constraints on learning -
@aaronbatista / Byron Yu). Showed low-D structure correlates w/ M1's ability to generate activity patterns (great follow-ups from@MattGolub_Neuro@jehosafet others).pic.twitter.com/cx8lpdo51s
-
3rd, our work (LFADS, Nat Methods w/
@SussilloDavid, others) - low-D/dynamic representations can be much more informative than observed, high-D data. E.g., much more closely tied to subjects' behaviors. -
Suggests that the low-D/dynamic representation is a more veridical view of network activity than observed neurons, i.e., we are extracting some property/state of the broader network.pic.twitter.com/q1cSsOyH0Z
-
Thanks for this, Chethan!
End of conversation
New conversation -
-
-
To add to
@chethan's great summary, in general there are many things that we can't understand about the brain without considering the population as a whole. This paper by Elsayed and Cunningham talks about this idea in more detail:https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5577566/ …Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Interesting question and answers. Any thoughts,
@ItsNeuronal@gamaleldinfe@neurotheory ? What do we learn by doing pca (or tca) on neural data? -
I think it's somewhat the wrong question, because PCA and related methods aren't designed to test specific hypotheses. Understanding data requires both exploratory and confirmatory analysis methods. I like John Tukey's classic book on this. PCA is mostly useful to explore...
-
Insofar as exploratory data analysis is incredibly useful to any scientific field, PCA is super useful for neuroscience. I worry asking for concrete insights/result misses that larger point.
End of conversation
New conversation -
-
-
rids our thought process about the brain of the 1:1 anatomic structure-to-function relationships classically taught previously. Instead, allows us to focus on the dynamic roles played by neurons in various regions depending on which circuits/contexts the experiment is exploring
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.