What if it had only been verified that the em's overall behavior perfectly corresponds to its biological template (i.e. without corresponding subparts down to your chosen ground level)?
Since whole brains are not repeatable, verifying behavioral isomorphism with a target would require a small enough target that its internal interactions were repeatable. (Then, having verified the isomorpmism, you tile it across the whole brain.)
What if e.g. groups of neurons could be perfectly (and more efficiently) simulated, using an algorithm which doesn't need to retain a "synapse" construct?
I would believe in this after someone had shown extremely high-fidelity simulation of synaptic compartments, then demonstrated the (computational) proposition that their high-level sim was equivalent.
Do you feel that some of the biological structural features on some level of granularity need to have clearly identifiable point-to-point counterparts in the algorithm?
No, but it's sufficient to establish causal isomorphism. At the most extreme level, if you can simulate out a synapse by quantum fields, then you are very confident in your ability to simulate it because you have a laws-of-physics-level understanding of the quantum fields and of the simulation of the quantum fields.
Since in any case, "verified surface correspondence" is a given (i.e. all em-implementations aren't differentiable from a black-box view)
Only in terms of very high-level abstractions being reproduced, since literal pointwise behavior is unlikely to be reproducible given thermal noise and quantum uncertainty. But it remains true that I expect any disturbance of the referent of "consciousness" to disturb the resulting agent's tendency to write philosophy papers about "consciousness". Note the high-level behavioral abstraction.
The combination of verified pointwise causal isomorphism of repeatable small parts, combined with surface behavioral equivalence on mundane levels of abstraction, is sufficient for me to relegate the alternative hypothesis to the world of 'not bothering to think about it any more'. There are no worlds of reasonable probability in which both tests are simultaneously and accidentally fooled in the process of constructing a technology honestly meant to produce high-fidelity uploads.
The combination of verified pointwise causal isomorphism of repeatable small parts, combined with surface behavioral equivalence on mundane levels of abstraction, is sufficient for me to relegate the alternative hypothesis to the world of 'not bothering to think about it any more'.
The kind of model which postulates that "a conscious em-algorithm must not only act like its corresponding human, under the hood it must also be structured like that human" would not likely stop at "... at least be structured like that human for, like, 9 orders ...
- Eliezer Yudkowsky, "Value is Fragile"
I had meant to try to write a long post for LessWrong on consciousness, but I'm getting stuck on it, partly because I'm not sure how well I know my audience here. So instead, I'm writing a short post, with my main purpose being just to informally poll the LessWrong community on one question: how sure are you that whole brain emulations would be conscious?
There's actually a fair amount of philosophical literature about issues in this vicinity; David Chalmers' paper "The Singularity: A Philosophical Analysis" has a good introduction to the debate in section 9, including some relevant terminology:
So, on the functionalist view, emulations would be conscious, while on the biological view, they would not be.
Personally, I think there are good arguments for the functionalist view, and the biological view seems problematic: "biological" is a fuzzy, high-level category that doesn't seem like it could be of any fundamental importance. So probably emulations will be conscious--but I'm not too sure of that. Consciousness confuses me a great deal, and seems to confuse other people a great deal, and because of that I'd caution against being too sure of much of anything about consciousness. I'm worried not so much that the biological view will turn out to be right, but that the truth might be some third option no one has thought of, which might or might not entail emulations are conscious.
Uncertainty about whether emulations would be conscious is potentially of great practical concern. I don't think it's much of an argument against uploading-as-life-extension; better to probably survive as an up than do nothing and die for sure. But it's worrisome if you think about the possibility, say, of an intended-to-be-Friendly AI deciding we'd all be better off if we were forcibly uploaded (or persuaded, using its superhuman intelligence, to "voluntarily" upload...) Uncertainty about whether emulations would be conscious also makes Robin Hanson's "em revolution" scenario less appealing.
For a long time, I've vaguely hoped that advances in neuroscience and cognitive science would lead to unraveling the problem of consciousness. Perhaps working on creating the first emulations would do the trick. But this is only a vague hope, I have no clear idea of how that could possibly happen. Another hope would be that if we can get all the other problems in Friendly AI right, we'll be able to trust the AI to solve consciousness for us. But with our present understanding of consciousness, can we really be sure that would be the case?
That leads me to my second question for the LessWrong community: is there anything we can do now to to get clearer on consciousness? Any way to hack away at the edges?