Building Smarter Artificial Intelligence By ... Shrinking The Body?
scientificblogging.com
scientificblogging.com
If a smaller brain is more intelligent than a bigger brain, I assume it is more complex (for instance, it has more interconnections).
In computer science though, it takes more processing power to simulate more complex representations of reality. Therefore, we would need more processing power to represent a small intelligent brain than a big stupid brain.
Software and our analog world are two very different beasts. Small and complex physical things sometimes take more processing power to represent than big simple ones.
The author doesn't seem to understand the causal link between intelligence, brain size, and body size. Smaller bodies don't make things smarter. Heavy organs have a survival cost. Animals with bigger brains only do better if the brains pay for themselves with survival-related intelligence. In species that don't generally get a payoff for figuring things out, the lighter and faster dummies breed better.
If you want to think clearly about AI, read Lesswrong.
Of course if you want AI that resembles humans, bodies might help. But why specialize too soon (creating AI is easier than creating AI that is like humans).
The problem is that "strong AI" (ie. an AI considered intelligent in the deepest sense of the term, with intelligence equal to or greater than that of a human) has pretty much run in to a dead end. Lots of approaches have been tried over the decades, and all have failed.
Strong AI research needs new ideas and new approaches. Making a close coupling between the "brain" (or "mind") and the body is one such relatively new approach. As such, I applaud it.
Like the approaches that came before, it also might lead up a blind alley. But at least they're trying something new and different, instead of banging their head on the same wall.
And, apart from the sheer novelty of the approach, there has been a lot of evidence that much of the "processing" that has been assumed to be done in the brain is actually done in the body.
Some relatively recent robotics research has been focused on creating robots without any kind of a "brain", but just with a "nervous system". These robots are able to move around and achieve goals. Here's some information on a conference where these sorts of robots are demonstrated:
http://ine-web.org/telluride-conference-2010/telluride-2010/...
These robots are a far cry from achieving the goal of strong AI, but they point to an alternative approach to AI that is not brain-centric, to coin a term.
I beg to differ with that.
Marvin Minsky's idea of having 5000 people creating a common sense database of sorts wasn't tried. (the closest is cyc, which only had 20 people most of the time).
Genetic programming has not been tried at all really.
Machine learning hasn't been seriously tried either (the closest efforts have been google goggles and google search maybe).
AI is serious and complicated engineering project. Billions are spent on creating skyscapers, tunnels, and airplanes. Yet the money spent on AI is orders of magnitude less even though the payoff is orders of magnitude more.
If I had to guess, I'd say AI would be easily achievable within a year using any of the above approaches if a trillion dollars were spent on it this year.
And the progress in the tools will eventually lead to more "intelligent" systems.
How has the name of AI been tarnished. If people believe you can spend a fraction of the money spent on a building to create human level intelligence then it's their 'intelligence' that is tarnished.
If your goal is create human level ai, than more 'intelligent' systems is not progress, we've had 'intelligent' tools since the first loom weaving machines were created hundreds of years ago.
One of the keys may be to be very selective about the kind and quantity of input and outputs the AI can have: the less it has to control itself, the more it can think.
Purely speculative, of course.
This has crap-all to do with a 'body'. This is such an obvious case of someone not defining their terms that it bugs the hell out of me. In the terms of the Dictionary from Anathem, it comes across as "bullshit" in sense 2.
We are still at the point where we don't have a whole enough picture to completely figure out the math of what makes a thing "think". We have ideas, hypothesis, and some theories. But its all very much in flux.
People quite often make the wrong analogy between the discovery of flight and conscious AI.
Before engineers knew about how air moving over a curved surface creates pressure differentials that provide lift, they tried to build flying machines by simulating wings. For decades, these people were laughed at by the common folk who believed human flight was impossible.
However, once it was discovered how air movement made pressures and it was possible to mathematically describe it, the approach changed. Then the first successes happened. Then there was a huge amount of innovation in the next 40-50 years before the innovation in plane and wing design leveled off.
Figuring out how to organize matter so that it "thinks" is still being figured out. We have some ideas that are well-supported by data. I feel the approach IBM is taking is among the best. Since we don't know the completely mathematical theory that applies to neuron networks to allow some to be conscious and others not, our best chance is to take as much real world data as we can and to build one in simulation. On the way, we discover all kinds of things we didn't know before.
In a way, it sounds like I'm countering my previous argument about how people trying to build wings failed. But, I'm not. It was in the process of attempting to re-create examples of structures that obviously could fly (birds) that we figured out the principles of how they flew.
Blue Brain probably won't become THE method of building AI in the future. However, its one of the first massive attempts to do the modern equivalent of building a wing from the atom up. Along the way, we will start really figuring out exactly what needs to happen to make one network of neurons conscious, and another one not conscious. Then we'll be able to take those mathematical algorithms and start playing with them.
Eventually we will get to a point where we can test an entity for consciousness. Once we can scientifically describe how consciousness works, we can devise tests for neuronal networks of any complexity to say if its actually conscious or not. This opens up the prospect of seeing if other animals on earth ranging from ants to dolphins are conscious in the same way that we are (or not).
It also gives us an actual true test for AI consciousness. If we see the same or similar results on a human that we do for an AI, its fairly obvious that if the AI says "I think, I am" then it probably is.
I've always had a problem with the Turing Test. Its too subjective. Just because you can fool 8 of 10 people, or even 10 of 10 people into thinking something typing to you is conscious doesn't mean it is. You CANNOT determine that without analyzing the structure of the mind that you're communicating with. If that mind is based on biology, you examine the network of neurons. If its based on software, you analyze the structure of the programming and the flow of data. Since the biological systems and software systems would both be using similar patterns of information to describe their structure and functioning, the same tests apply to both.
Sorry to get off on a rant. A "body" has shit all nothing to do with "making a better AI".
So you're saying Rodney Brooks's (previously head of AI at MIT) behavioural ai is wrong. Also, without 'sensory' data from a 'body' machine learning approaches can't work either.
"""Eventually we will get to a point where we can test an entity for consciousness. Once we can scientifically describe how consciousness works, we can devise tests for neuronal networks of any complexity to say if its actually conscious or not """
Are you sure there is such a thing as conciousness? Even if there is, I'd bet we get to AI before we can test for conciousness.