Monday, May 28, 2012

Human intelligence is overrated

People often wonder when computers will become as intelligent as humans. This question assumes that being as intelligent as a human is a worthy goal for an artificial system. But really, are humans all that smart? I don't think so.

Humans are quite stupid in many ways, compared to computers. Let's start with the most obvious: they can't count. Ask a human to raise 3425 to the power of 542 and watch them sit there for hours trying to work it out. Ridiculous. The same goes for a number of other trivial tasks, such as calculating the average age in a population of 300 million. Shouldn't take more than a couple of seconds. Unless you are a human, in which case it'll probably take you years, and even then you would have made a number of errors.

Humans have almost no memory, either. Ask a human to give you the correct name and current address for a randomly chosen personal registration number (or national insurance number, or whatever the equivalent is in your country). Even if they have all the information in whatever format they prefer, it will still take them at least several seconds - and most humans would not even know where to get the information. Or ask a human to produce 100 URLs to websites talking about artificial intelligence, or even a complete list of everything that happened to them yesterday. Humans talk about "goldfish memory", but from the perspective of a computer the human and the goldfish aren't that far apart, capability-wise.

At this point, many humans will be protesting wildly and saying we are being terribly unfair to them. We are only choosing tasks that computers excel at, and ignoring those where humans have an advantage, such as motor control and pattern recognition.

Right. Computers can land a jet plane and fly an helicopter. In fact, almost any computer can do it, if you just load the right software. Very few humans can land a jet plane and even fewer can fly an helicopter. Sometimes humans fail spectacularly at these tasks. (It's hard to understand why anyone would want to be in a plane flown by a human, now that there are alternatives.) Computers can drive regular cars on-road and off-road, obeying all traffic regulations. There are many humans that can't even do that.

Speaking of pattern recognition, it's true that humans can recognise the faces of their friends with quite high accuracy. But then, humans only have a couple of hundred friends, at most. The face recognition software that Facebook uses can tell the faces of millions of people apart. Other pattern recognition algorithms can successfully match a scan of a human thumb to the right fingerprint in a database of millions.

Now let's take another activity that humans should be good at: game-playing. Games were invented by humans in order to entertain themselves, and as humans seem to find it entertaining to exercise their learning, motor and reasoning capabilities, games should be perfectly tailored to human intelligence. Humans should excel at game-playing, right? Well, no. Since 1997, the world Chess champion has been a computer. The situation is even more extreme for Checkers: computers have computed how to play the game perfectly, meaning that it is mathematically impossible to win over a computer unless you follow the very same strategy. Moving from board games to video games, we have a similar situation. There are AI players that play most levels of Super Mario Bros better than any known human, and computers kick human ass in first-person shooters like Unreal Tournament because humans have such lousy reaction times. Mind you, all these are games that were designed by humans for humans. It would be very easy to invent games that were so complicated that only computers could play them; computers could even invent such games automatically.

Other things that have been cited as pinnacles of human achievement are tying shoelaces and self-reproduction. But tying shoelaces is sort of pointless, it's an obsolete technology even for humans; why would you need shoelaces if you're a robot? And humans don't really know how to reproduce themselves. They know how to have a sex, which is quite a different thing and really sort of easy. The actual reproduction is down to various biochemical processes that humans don't even understand yet, much less can they replicate them.

Now, some humans reading this will be all up in arms, and accusing me of dishonesty and sophistry in trying to degrade human intelligence. I'm giving all these examples of computers being good (and humans bad) at performing very specific tasks and solving very specific problems, when the hallmark of real intelligence is to be able to perform well in a large variety of situations. According to this definition, the human would say, computers are not very intelligent: a chess-playing program cannot land a jet plane, and a face recognition program cannot play Super Mario Bros, nor can it do exponentiation.

This is true. But the argument goes both ways here as well: take an arbitrary human (such as yourself, if you happen to be human) and try placing this human in the cockpit of a landing jet plane, in a semiconductor factory, in the oval office of the White House, in the kitchen of a gourmet restaurant, on a horseback in Siberia, or equipped with only a spear in the middle of the Amazonas jungle. There are humans that have been programmed to do well in each of these situations, but it is very unlikely that the human you were thinking of (perhaps yourself) would know what to do in more than at most one of these situations.

So, compared to humans, computers seem to be doing quite well indeed. If man is the measure of all things, artificial intelligence research should be declared a success already. But that would be like saying that the development of the car would have been finished when cars first became as fast and reliable as horse-drawn carriages. As we know now, it was certainly possible to do better than that.

It will take some time, and much hard work, before we can say that computers are truly generally intelligent. But, to be fair, we should be careful with ascribing general intelligence to humans as well - humans are just not that smart. In particular we should be very careful with using human intelligence as a measuring yard for machine intelligence, and we should stop asking when computers will become as intelligent as humans. It seems that man is the measure in this particular case simply because we didn't know any other type of intelligence when we started thinking about the problem of artificial intelligence. Hopefully that will change soon.


anne.ominous said...

This article was a waste of time.

You are confusing issues in exactly the same way that people did around the turn of the 20th century (i.e., 110+ years ago):

You are confusing intelligence with computing power. They are not even remotely the same things. Thus the big problem of AI, which nobody has managed to come close to solving yet: how to take all that computing power and turn it into INTELLIGENCE. Which means being able to THINK about all that data, not just process it.

And no, before you go there: Watson ain't it. It's just a supercomputer doing many clever CALCULATIONS per second, according to rules that intelligent HUMANS wrote. It has never even come close to having a thought of its own.

I have been watching the field of AI as a hobby for over 30 years now. And believe me, you are WAY off. The researchers are barely closer to it now than they were then.

ErezSh said...

I found the peace very amusing, but it's also incorrect.

Humans have several general abilities that computers don't yet have.

1. Learning. Humans can learn to do complex tasks and figure out new situations on their own. Put a city-human and a city-computer in a jungle, and we'll see who makes it out of there.

2. Contextual search. We can execute big complex queries in our brains, with many obscure and interconnected conditions, and get the results almost instantly. Computers today fumble to perform a fraction of such complexity.

3. Math. No, I don't mean arithmetic. I mean Math. What important mathematical contributions can be attributed to unaided software? I'm guessing none, there were always humans telling it exactly what to look for.

4. Recaptcha?

I could go on for quite a while, I suppose. Unfortunately, computers still have a long way to go.

Unknown said...

That's a nice point you make. You definitely lowered my appreciation of the general intelligence of humans, which is indeed the number one counter argument when someone asks me: are computers smarter than humans?

I also agree that we should use other measures of intelligence than resemblance to human intelligence. However, I think there's still a valuable thing to learn from humans, and that's learning or adapting. Like you made clear, a computer with the right software can perform most or possibly any task better than a human can. The software however is still written by human programmers. While AI can be excellent at execution of any given procedure, it is often a human that devised this procedure.

Now, I know that it would be easy to write another paragraph that shows that computers are excellent learners and can perform very well at problems of which the solution is not explicitly given at the time of programming. Still I think that this learning is not at all generally applicable. The learning of any task by a computer should be closely designed and monitored by a human with good results, and mostly does not exceed expectations. Smart humans can easily switch context and perform reasonably well in unexpected tasks. Given a task or a problem that they are not 'programmed' for they can use a general problem-solving procedure to solve it.

While I agree with most of your writing, I think AI still needs to develop a more general and self-initiated way of learning to compete with humans on all forms of intelligence. Of course I do admit that we lose on most :)

Anonymous said...

Computer failures happen a lot in planes, and computers are still too unreliable to be allowed in landing or takeoff by the ACC unless the pilot is out of commission and there is no alternative.
Also, the Google car is the most advanced and is still incapable of obeying all traffic laws...
And people aren't best at math because our minds are so abstracted so we can focus on many tasks at once.
Facebook's face-recognition algorithm only works with the -exact- same image or the exact same face-position at that age.
Humans can recognize similar faces from multiple directions/angles.
No AI has been capable of beating a Human at Go, or even doing well at the game. Go's older than Chess and seen as the ultimate in strategy games.

Computers can't even solve problems yet. They're further from being capable of deciding what the problem is. And even further yet from being able to determine the best solution.

Once a computer can analyze a situation, find the problems to solve in the situation, and then write programs to solve said problems, they will be as intelligent as Humans.

immensetalent said...

the only thing you neglect to mention is that humans design and code computers. I'm pretty sure that computers can run code, but can't write it.


Nathan said...

I'm assuming these are the sorts of arguments that robots will make to explain why they should be in charge.

Jonathan said...

Its just mortals evolving nothing to read into you would understand or need to: