Nicole Rust on Nostr: Is an "algorithm" - like a line attractor in a (recurrent) neural network - a ...
Is an "algorithm" - like a line attractor in a (recurrent) neural network - a mechanism?
Many great papers about the brain describe the "algorithms" by which it computes. Among them are those that model it as complex dynamical system that operates by settling into different states called attractors; one example is Mante et al 2013 who used this approach to understand how prefrontal cortex contributes to decision making.
https://www.nature.com/articles/nature12742
Is that explanation mechanistic? To me: obviously. But I've just learned that some argue it's not. What's the logic? It begins by explaining that "mechanism" implies "causality" and it proceeds to question whether these types of explanations are causal.
With regard to that study:
https://www.journals.uchicago.edu/doi/abs/10.1093/bjps/axw034?journalCode=bjps
"One strike against the mechanistic interpretation is that the explanation offered by the dynamical model of the PFC is not a constitutive one. That is to say, it is not an explanation of how a global phenomenon— computation—comes about because of the activities of some network components, whether spatially localized or not. Instead, the dynamical properties of the network, which do the job of explaining the computational phenomena, are themselves global- or population-level properties of the network.28
...
Another important point is that the model gives us no information about how we might make changes to the network in order to affect changes to its information-processing properties. It does not tell us which connections would have to be rearranged in order to make the computation no longer context-dependent, for example.
...
This brings me to say more about the second option, which I endorse: that the model offers some kind of non-causal explanation. We can think of the dynamical model as offering an illuminating perspective on the network."
I have problems with this interpretation. In my own mind, algorithmic explanations are mechanistic. Curious to hear your thoughts!
BTW: My dive into this discussion was prompted by this paper:
https://www.nature.com/articles/s41583-023-00778-7.epdf?sharing_token=LlREMtfLsEuMP2cIGP8-CNRgN0jAjWel9jnR3ZoTv0OJW18pxUOzYI6WS9QWYKLbrMfHUnYQTzLgDtF-GegxI1N6TrMJPNhyio4bQPoB5cqsg7EyaZJfZng31XWQB5hPeXriZftbC-ae313h2IZqMg9bAlGExqGCnik85rG07zQ%3D
where the authors don't take a stance on algorithm per se but advocate for the field to arrive at some clarity around what "mechanism" means.
Many great papers about the brain describe the "algorithms" by which it computes. Among them are those that model it as complex dynamical system that operates by settling into different states called attractors; one example is Mante et al 2013 who used this approach to understand how prefrontal cortex contributes to decision making.
https://www.nature.com/articles/nature12742
Is that explanation mechanistic? To me: obviously. But I've just learned that some argue it's not. What's the logic? It begins by explaining that "mechanism" implies "causality" and it proceeds to question whether these types of explanations are causal.
With regard to that study:
https://www.journals.uchicago.edu/doi/abs/10.1093/bjps/axw034?journalCode=bjps
"One strike against the mechanistic interpretation is that the explanation offered by the dynamical model of the PFC is not a constitutive one. That is to say, it is not an explanation of how a global phenomenon— computation—comes about because of the activities of some network components, whether spatially localized or not. Instead, the dynamical properties of the network, which do the job of explaining the computational phenomena, are themselves global- or population-level properties of the network.28
...
Another important point is that the model gives us no information about how we might make changes to the network in order to affect changes to its information-processing properties. It does not tell us which connections would have to be rearranged in order to make the computation no longer context-dependent, for example.
...
This brings me to say more about the second option, which I endorse: that the model offers some kind of non-causal explanation. We can think of the dynamical model as offering an illuminating perspective on the network."
I have problems with this interpretation. In my own mind, algorithmic explanations are mechanistic. Curious to hear your thoughts!
BTW: My dive into this discussion was prompted by this paper:
https://www.nature.com/articles/s41583-023-00778-7.epdf?sharing_token=LlREMtfLsEuMP2cIGP8-CNRgN0jAjWel9jnR3ZoTv0OJW18pxUOzYI6WS9QWYKLbrMfHUnYQTzLgDtF-GegxI1N6TrMJPNhyio4bQPoB5cqsg7EyaZJfZng31XWQB5hPeXriZftbC-ae313h2IZqMg9bAlGExqGCnik85rG07zQ%3D
where the authors don't take a stance on algorithm per se but advocate for the field to arrive at some clarity around what "mechanism" means.