Comment by kmoser
2 months ago
A Turing machine operates serially on a fixed set of instructions. A human brain operates in parallel on inputs that are constantly changing. The underlying mechanism is completely different. The human brain is far, far more than a mere computation device.
Efforts to reproduce a human brain in a computer are currently at the level of a cargo cult: we're simulating the mechanical operations, without a deep understanding of the underlying processes which are just as important. I'm not saying we won't get better at it, but so far we're nowhere near producing a brain in a computer.
Any Turing complete computational device is computationally equivalent to any other, irrespective of whether it carries out the computation serially or in parallel or on a fixed set of instructions or an infinite set.
Unless you can demonstrate that the human brain can compute a function - any function - that exceeds the Turing computable, there is no evidence to even suggest it is possible for a brain not to be computationally equivalent to a computer.
I'm not saying it's impossible. I'm saying it's highly improbable to ever come close to fruition, since if you want to be completely accurate in your computations, you'd have to model the physical effects of every single atom in the brain and how they affect each other, in realtime (physically, chemically, electrically, etc.). Not only would that be computationally expensive, but you'd also need to know the complete set of rules for doing so, which includes how things interact at the quantum level.
Unless there is unknown physics going on, we have the existence proof for the possibility of a compact computational device capable of accurately replicating the thought patterns of a human in the form of every single human brain. That, to me, makes it highly improbably that we won't be able to replicate it at some point, not least because we know that human brains remain functional even with significant variations even with substantial amounts of damage.
So while it might well be we will need new architectures - maybe both software and hardware, it seems highly unlikely we won't be able to.
There's also as of yet no basis for presuming we need to be "complete accurate" or need to model the physical effects with much precision. If anything, what we've seen consistently over decades of AI research is that we've gotten far better results by ditching the idea that we need to know and model how brains work, and instead statistically modelling outputs.