Hawking: A.I. 'worst event in the history of civilization'

Hawking: A.I. 'worst event in the history of civilization' Nov 17, 2017 1:09:09 GMT

Quote

Post by general313 on Nov 17, 2017 1:09:09 GMT

Nov 16, 2017 14:33:56 GMT faustus5 said:

Nov 15, 2017 20:35:02 GMT general313 said:

But there does seem to be a repetition of that tendency, oft-repeated throughout the history of science, that the closer we look, the less difference we see between the living and the non-living.

When you are looking at the lowest levels, watching protein formation or the biochemistry occurring at a single synapse, yes: completely mechanical processes.

So you think there is something special that happens at a higher level in the brain that is beyond the reach of machine learning designers? Some kind of emergent property? As I said before, given the history of observers imaging some fundamental and special difference between the biological and non-biological, I won't bet on it.

As neural scientists probe deeper into how human vision works, they are discovering more about how the brain processes images for recognition, and it bears a striking resemblance to the neural networks used in machine learning.

Well, you have the resemblance kind of backwards there. Neural networks were inspired by what scientists saw going on in the brain, so the resemblance is hardly "striking" since it was by design!

It doesn't matter where the inspiration came from. If we can create artificial neural networks that exhibit the same behavior as brain structures it takes away some of the mystery of how the brain works. That the behavior can be replicated with inorganic machinery suggests that there's nothing special about biological brain cells. It's simply a factor of how we evolved,
just as we locomote via muscles and bones, and not with wheels.

Learning and memory retention in human brains and neural networks work the same way, by adjusting interconnection weights (synapses in the case of humans and layer node coefficients in a neural network).

Only at the most superficial levels is it the "same". Neural networks fail to capture the complexity going on inside of individual neurons, which are living creatures competing with each other for resources, and also leave out the important role of the biochemical soup neurons live in, which influences their firing patterns in important ways. I should add that there are also models which posit that considerable information processing/computing activity goes on even within individual neurons, not just in between them and in their networks.

Individual neurons functionally are not very different from neural network nodes. A brain neuron can have thousands of interconnects (dentritic connections) but so can neural network nodes. Would you have links to these models with "considerable information processing happening within a neuron"? It seems to me that the interior of a neuron is mostly the cellular machinery typical of other cells for metabolism (managing energy to operate).

You dismiss "simple mechanical circuits", but brains and smart computers derive their power in the organization of very large numbers of simple elements.

The mechanical circuits built by humans pale in complexity to their counterparts in the brain. Even the most complicated are extremely simple and crude when compared to biological systems. Remember, my motto in all this is that you have to take functionalism seriously, which means applying it all the way down and not stopping until you've instantiated every relevant cause and effect in your alternate medium. Only then can you say you've created a model that is truly substrate-neutral.

That is true so far, but that is mainly a question of the number of neurons that are in the human brain (100 billion or so I believe). Admittedly we are still in the infancy of machine learning, but I think there is every reason to expect great advances in our ability to make hardware powerful to rival the capabilities of the human brain in the not too distant future. Hence the concern of the main topic of this thread.