4 ms·

Looks like the beginning of Permutation City[1] - won't be long now (relatively) until we won't be able to observe a difference between real c.elegans and simul

Looks like the beginning of Permutation City[1] - won't be long now (relatively) until we won't be able to observe a difference between real c.elegans and simulated c.elegans, simulate its environment, simulate other organisms, etc. Will these worms be _alive_? What happens when computers become advanced enough to simulate humans? What if they already are? o_O 1. http://en.wikipedia.org/wiki/Permutation_City http://en.wikipedia.org/wiki/Permutation_City

drostie 14y ago
Well, hold on a second. Even if we know the worm's neural network and how it's connected, that doesn't necessarily mean that a computer simulation of it will do exactly what the worm does. The problem is that we're not doing an atom-by-atom simulation of the worm but rather we're hoping we've extracted the important logical features of how the worm works, so that the thing will run on modern computers. This is actually the more exciting part of the research. Serious ethical questions aside[1], imagine if you could simulate a human's neural network. Probably your first simulation would seem unbelievably stupid. Perhaps you find out that you didn't account for the ways certain neurotransmitter concentrations will "leak" data from one synapse to a nearby one, without which the system becomes radically disconnected because evolution was lazy and connected them without a specific wire. (I don't know; I'm being hypothetical.) So now you get to introduce some sort of adjacency matrix which manages which neurons are "next to" each other and receive these "secondary signals." Then it seems to be able to learn language, but it still can't balance in the world, and working it out, you find out that there is a big failure in the motor regions because they only work when the right signal propagation delays are introduced, and you were propagating them all instantaneously, and so on. In the distant past, we had hoped that chess was so complicated that it would only be solved with some great insight into human understanding -- but instead it was solved with brute force. This is one of the first cases where I see that the brute force might be finally able to give us a test model by which we might better understand understanding itself. The only worrying bit is the neural nets themselves. Neural nets are notoriously difficult to interpret and understand. Even the calculus-based approach of "I'm going to make a tiny tweak to the network and see how it changes the output" offers only a little enlightenment. [1] I do think ethical questions about killing artificial consciousnesses deserve discussion time; I just don't have much to spare at this moment and it's kinda tangential.

drostie 14y ago

Well, hold on a second. Even if we know the worm's neural network and how it's connected, that doesn't necessarily mean that a computer simulation of it will do exactly what the worm does. The problem is that we're not doing an atom-by-atom simulation of the worm but rather we're hoping we've extracted the important logical features of how the worm works, so that the thing will run on modern computers. This is actually the more exciting part of the research. Serious ethical questions aside[1], imagine if you could simulate a human's neural network. Probably your first simulation would seem unbelievably stupid. Perhaps you find out that you didn't account for the ways certain neurotransmitter concentrations will "leak" data from one synapse to a nearby one, without which the system becomes radically disconnected because evolution was lazy and connected them without a specific wire. (I don't know; I'm being hypothetical.) So now you get to introduce some sort of adjacency matrix which manages which neurons are "next to" each other and receive these "secondary signals." Then it seems to be able to learn language, but it still can't balance in the world, and working it out, you find out that there is a big failure in the motor regions because they only work when the right signal propagation delays are introduced, and you were propagating them all instantaneously, and so on. In the distant past, we had hoped that chess was so complicated that it would only be solved with some great insight into human understanding -- but instead it was solved with brute force. This is one of the first cases where I see that the brute force might be finally able to give us a test model by which we might better understand understanding itself. The only worrying bit is the neural nets themselves. Neural nets are notoriously difficult to interpret and understand. Even the calculus-based approach of "I'm going to make a tiny tweak to the network and see how it changes the output" offers only a little enlightenment. [1] I do think ethical questions about killing artificial consciousnesses deserve discussion time; I just don't have much to spare at this moment and it's kinda tangential.