Brian – I saw Randy’s Science paper and (probably like yourself) was very surprised

I understand that they made the mistake of only considering single-layer networks when pouncing on perceptrons; if they had considered multi-layered networks they would have seen that things like XOR are possible. Linearity and analog systems notwithstanding, I can say with the hindsight of a huge generational gap that it just seems silly to me that they didn’t consider multi-layered networks.

A consistent thread in your comment is that some differences are merely « implementational » or « architectural » details, and thus are actually unimportant or otherwise superficial. IMO, that attitude is scientifically dangerous (how can you know for sure?) and *very* premature (when we have an artificially intelligent digital computer, I’ll be convinced).

Just as the brain only has hardware (as you said, there is no software that is the mind running on top), the only thing that counts when programming a mind is the software. The high-level software need not be concerned with memory registers when representing knowledge, and « pointers » can be implemented on a system that uses only non-volatile memory.

I think the only real problem here is whether or not digital computers can simulate the continuous nature of the brain. If it is the case that a discrete state machine is not hindered by this, then the brain’s architecture with all of the intricacies of neuronal activity can be implemented to the fullest extent with no other problem (although we’d of course want to abstract away as much complexity as possible). However, if digital computers cannot simulate continuous structures with sufficient robustness, then I think AI would have to start putting more research into analog circuits. But I don’t think we currently have enough evidence yet to make the case for either.

So yes, brains and PCs have different architectures, but that doesn’t mean you necessarily cannot implement a mind on a computer. […]