Graham Findlay

PhD student | Neuroscience of sleep and consciousness | University of Wisconsin-Madison

[Graham Findlay @grahambfindlay]

@wmayner @davidchalmers42 "causal connections seem to be grounded in a combination of the intrinsic powers of neurons and extrinsic synaptic processes connecting neurons. it's also not obvious how one would formalize this constraint..." Exactly what the next version aims to achieve. 🤞

[Graham Findlay @grahambfindlay]

@wmayner @davidchalmers42 This is a very good summary. A word of caution about that preprint, though: it is very unfinished. It has already changed a lot. That first pass was, in Dave's words, "a very strong condition that might rule out all sorts of systems, arguably including neural systems." 👇

[Graham Findlay @grahambfindlay]

@davidchalmers42 Not simpler than what you already surmised:
1) Many extrinsically appealing blackboxings do not comply with the postulates.
2) Systems of compliant macro units must still themselves satisfy the postulates.
I think (2) is the deeper issue.

Expanded:
https://t.co/4MrDBNYAgh

[Will Mayner @wmayner]

Hi Dave, thanks, you raise a good question.

The argument for IIT 3.0 is basically that with a von Neumann architecture, you'd need to either treat the CPU as background or as a constituent of one or more of the units of the PA being implemented, and:

(1) Treating the CPU as a background condition doesn't work, because the IIT analysis stipulates that the candidate system is causally conditioned on external units in their current state. This effectively screens off the CPU's causal powers, so that the "relevant strong conditionals", to use your language, are broken. Roughly speaking, the CPU can't simultaneously be doing the causal work of all the 1000 units once we condition on its current state. (Personally, this conforms to my intuition that in order to place the CPU in the background while retaining the function of the 1000 units, there must be some kind of fishy spatiotemporal gymnastics going on with the micro-to-macro mapping.)

(2) Including the CPU as a constituent of a single macro unit doesn't work either, for similar reasons of causal screening. When the IIT analysis considers subsets of the system that don't include that unit, they are likewise causally conditioned on the CPU.

(3) Counting the CPU as a constituent of each macro unit is ruled out, as you suspected, by the requirement that black boxes can't share common micro-constituents. In the black-boxing paper, this is motivated by appeal to exclusion (the macro units must be definite), but I think there are also independent reasons to impose some such restriction (your paper "Does a rock implement every FSA" and Putnam's work are cited in the section of the black-boxing paper that describes the non-overlap condition). The paper where this is spelled out is here: https://t.co/yogGRuvaJP

For IIT 4.0, as others have already pointed out, there is now an even more restrictive condition on mappings (roughly, macro units must be maxima themselves, "maximally-irreducible within"). There are some sentences about the motivation in the 4.0 paper. There's also this preprint with details: https://t.co/bCiFW7ix3R

The justification for this stronger constraint is perhaps more subtle than in the IIT 3.0 case and more connected with the underlying ontology of IIT, so I'll leave that to the paper that @grahambfindlay mentioned, hopefully out soon—though I think it's still in the spirit of using IIT's principles and postulates to pick out the right notion of implementation & physically-instantiated causal power.

Hope this helps, happy to discuss more.

@LAlbantakis @mttgrs @theBjornErik @NaoTsuchiya @JohannesKleiner

[Graham Findlay @grahambfindlay]

@davidchalmers42 I know that this does not answer your most important questions. The result is important, but most important is why it follows from the theory's core principles. Much to be said. Would be honored to discuss. But wanted to quickly provide this context. Public preprints soon 🤞

[Graham Findlay @grahambfindlay]

@davidchalmers42 Importantly, nothing changed qualitatively about the results from 3.0 to 4.0. The restrictions on the macro units have changed, but restrictions were always there. The core argument is the same. Actually, the result was harder to prove using 4.0 than with 3.0.

[Kenneth D Harris @kennethd_harris]

Interesting discussion on neurophysiology data standards going on. Here are some thoughts. What follows is not a criticism of the NWB format, which has hugely benefitted our field. As part of the team that designed NWB1.0, I can even share some credit for it.+

[Larissa Albantakis @LAlbantakis]

New preprint. The magnum opus on Free Will and IIT. Features: the first proper account of IIT's ontological commitments and how those imply that you have free will. (Spoiler: you exist, your neurons do not truly exits, only what exists can cause).
https://t.co/WmbljqYTqG