Will Artificial Intelligence Make Humans Extinct?

User avatar
neeneko
Posts: 1092
Joined: Fri Jun 02, 2017 9:08 am

Re: Will Artificial Intelligence Make Humans Extinct?

#26

Post by neeneko » Fri Oct 20, 2017 11:59 am

RTH10260 wrote:
Fri Oct 20, 2017 11:51 am
Dubbed AlphaGo Zero, the Artificial Intelligence (AI) system learnt by itself, within days, to master the ancient Chinese board game known as "Go"—said to be the most complex two-person challenge ever invented.
As someone who has spent a lot of my career working in AI, this actually really saddens me.

Over the last decade or two, 'machine learning' has become the big go-to AI technique, producing all sorts of useful results. It has become the industry, if not 'standard', at least 'buzzword' for how to solve pretty much any problem.

It has a huge downside though. Researchers learn almost nothing from it. From an industrial perspective such systems tend to be useful, you figure out what inputs to give it, optimize where you can, throw as much CPU and memory at it as you can afford, and magic comes out the other side. But ask people who work in the field how it came to its conclusions or what models they applied and you tend to get blank stares. People are learning to set up black boxes, but research into how things actually work or what underlying patterns exist has really been falling out of favor.



User avatar
RoadScholar
Posts: 6619
Joined: Wed Jan 26, 2011 10:25 am
Location: Baltimore
Occupation: Historic Restoration Woodworker
Contact:

Re: Will Artificial Intelligence Make Humans Extinct?

#27

Post by RoadScholar » Fri Oct 20, 2017 12:27 pm

I read about the same situation arising from self-designing electronics. A person designs an electronic module made to produce result X as best as possible. The machine is told to make multiple changes to the circuit and look at the results. It then takes the two best designs and uses them as seeds for the next batch of "mutations." The only goal is achieving the best result.

They ran the program for quite a while. At the end, the machine had built a circuit that performed spectacularly... but when the people examined what the computer had designed, the didn't understand it at all. It simply didn't make sense to them, but somehow it worked. They reported that the experiment was disappointing in that they didn't really learn anything about designing better circuits.


The bitterest truth is healthier than the sweetest lie.
X3

User avatar
neeneko
Posts: 1092
Joined: Fri Jun 02, 2017 9:08 am

Re: Will Artificial Intelligence Make Humans Extinct?

#28

Post by neeneko » Fri Oct 20, 2017 4:22 pm

RoadScholar wrote:
Fri Oct 20, 2017 12:27 pm
They ran the program for quite a while. At the end, the machine had built a circuit that performed spectacularly... but when the people examined what the computer had designed, the didn't understand it at all. It simply didn't make sense to them, but somehow it worked. They reported that the experiment was disappointing in that they didn't really learn anything about designing better circuits.
I have no idea if it is still topical, but I can recall quite a bit of debate around issues of 'smart' optimization and compilers. They could produce machine code that no programmer was going to be able to make sense of, and while it was faster it was hell to audit and could behave in problematic ways when it hit strange situations. This could be a real problem when it came to validating and underwriting code that _HAD_ to behave certain ways.

Which is a concern I hear regarding machine learning. When it is just advertising or marketing or search mistakes are harmless. When money or lives are on the line 'I don't know how it works, it just does' starts becoming a problem. Last time I talked to someone in underwriting they kept to explicit equations for just this reason. Sure they were less clever than machine learning, but you could pick them apart and explain why you got the results you did.



Post Reply

Return to “Computers & Internet”