Hacker News Reader

Comment by richardwhiuk | Hacker News Reader

AI is repeatedly redefined to exclude things that we can do, that were previously considered AI. Examples include image and character recognition, speech recognition, machine translation.

Each time we say 'oh well that's not actually the machine thinking it's just doing a search of a database / applying statistics / guessing' so it's not AI. Except we don't really know what it would mean for a machine to think, or understand text.

View on HN

I thought AI was about one algorithm or something like that could do general thinking like a human. Its like you recognize images, characters, speech and then? What will you do with that?

Of course you could argue a child can detect a cat is cat only because someone taught the child so. But the point is a human baby left in the jungle right after it was born without any outside help can do a lot of intelligent things. Can a machine do the same?

"I thought AI was about one algorithm or something like that could do general thinking like a human."

I think that was the view until the early 90s or so (although this may be popular again - AI research does seem rather cyclical) - encode a sufficiently rich database of "knowledge" fire up your "inference engine" and general intelligence results. What "knowledge" you need to bootstrap an intelligence, how to represent it and what "inference engine" to use being the source of much debate (and even more funding). This didn't really deliver an awful lot - so things like multi-agent systems and neural network became popular (or in the case of neural networks, became popular again).

You are very right. That's why I divide the world between Cognitive Psychology, which is "How intelligence really works" and Artificial Intelligence which is "Replicating and improving intelligence with machines." The former is a very long drawn out field. (The research projects that the professor is talking about have been going on for 20+ years.) The latter is difficult. So far they can only tackle "Making useful things that don't represent how humans do it" or "Modeling how humans do things in a way too simple to be useful".

The gaps between those are closing, but it seems like the field is actually pushing both boundaries in the opposite direction. We're making very useful things that look less and less human. (Think Google search) We're also understanding the human mind much better (Think connectionist models) but it's still hard to get them to produce real things. This isn't an awful thing. The field is still much better off than 20-40 years ago when it hyped universal general intelligence that could also be useful.

I would say that these algorithms display intelligence to the extent of their input and output abilities.

You won't get what Hofstadter considers as "real" AI (artificial general intelligence, or AGI) without physical robots who need to display survival instinct and skills in order to keep on working.

I'm not sure I want it to happen, to be honest, but I'm sure it will.

They don't have to mimic each and every quirk of humans to do so.

For example, human language is ambiguous, slow to produce and to understand. Packet-based networking and binary protocols are far more efficient. The same goes for our cognitive biases, even though some of them may be useful to survive, counterintuitively. A fast and energy efficient algorithm that works most of the time can be better than an exact one that's more expensive. See bloom filters, for example.

"without physical robots who need to display survival instinct and skills in order to keep on working."

That's "situated cognition" - the idea that knowing and doing are intrinsically linked:

http://en.wikipedia.org/wiki/Situated_cognition

There is no reason a robot body is necessary for intelligence. It might even hinder it as the real world is much more messy and complicated than cleaned up and relevant data fed to machine learning algorithms.

You could probably bootstrap it in a virtual world too, that's true. I don't know if the skill acquired there would easily translate to the mechanical world, though.

Me personaly, I define AI as a system that can be fed non-biased input and learn from that input to derive conclusions that we as humans can relate too.

So feeding a system lots of cat pictures (ie biased-input) to teach it what a cat is for me is not AI. But a system which you feed in lots of random pictures and it learns by itself what a cat is, that would be AI, least for me.

What will be really interesting is a system which you can feed in all the childrens section and see what comes out at the end, would be most insightful into how we teach children and what we teach them. So a completely different area of AI use from that alone - learning how to learn better.

>>But a system which you feed in lots of random pictures and it learns by itself what a cat is, that would be AI, least for me.

A very good definition. Also adding to your point. Lets say a machine is fed with a billion pictures. Can the machine automatically categorize them by reading through them? Doesn't matter if it refers to a cat as some alphanumeric name 'ab12er' or a dog as 'p09iuy'.

But it should be able to categorize them. Then it should be able to read some encyclopedia or some source of information and study the behaviors of 'ab12er' and 'p09iuy'. Or the opposite, see 'ab12er' and 'p09iuy' and recognize them and describe them what they are.

Thing is though by giving a encyclopedia you are feeding in a bias thru a form of base reference. Thinking it thru a system could learn by itelf what a cat was, but it would not know it by that name. With that the creation of AI could be truly artifical and intellegent, but in a form we do not comprehend or even understand, yet it could be in its own right intellegent.

Seems to me you are defining AI as a generalized categorization algorithm. This sounds very narrow as it doesn't take into account any actual intelligent behavior. For example, I believe constructing tools in order to accomplish a specific goal (eg, constructing a net in order to catch fish) to be intelligent behavior. The categorization AI you describe would be incapable of this. Though it does sound like it would be a necessary prerequisite.

You're describing supervised vs unsupervised learning. We have both systems, one isn't really defined more as AI than the other.

Good to know. Though thinking it thru, any unsupervised learning would have to have some reference points, be them hard codeded into the core system to ina form of supervised learnings. Otherwise the AI could recognise a cat but would know it by a completely different name and unless it had been given a good description of a cat to associate with such image forms or a picture with a label. Then it would never know what a cat was in the way we know them. Which may be a good or a bad thing. But if the AI referers to cats as 478912's then we would not know what it was on about and whilst it may be intellegent, it would be in a way that we would be unable to understand and relate too. Ironicly i suspect if you had a top end AI system and asked it what defined AI it may very well come back with the answer 42, which many would even understand, though no comprehend.

Fun feild of work still, I bet.

Well, you can show it a photo of a cat and ask it "what's this?", and it'll say "this is a 478912". That's how clustering works. Obviously it's not going to just speak to you, but there are ways of extracting the categories out of it (obviously, otherwise why run it if you can't read the results?)!

This is called Deep Learning. It can automatically form clusters from unlabelled training data and that is why there is a lot of excitement around it right now.

That's called unsupervised learning and it already exists. Supervised learning is more practical though. I'm not sure why you distinguish them because they are both AI and sometimes even the same algorithms.

For that to happen the AI needs to be pointed to what is a cat.

Children know cats by mimicking adults who point out the objects to them.

> machine translation

One look at the Play store and its machine translated descriptions shows how far we are from anything usable.

I think it shows how far we've come. Remember altavista's machine translation from babelfish?

Look at how far we've climbed this tree, we'll make it to the moon soon.

AltaVista/Yahoo Babelfish was based on SYSTRAN, as were earlier Google Translate versions! http://en.wikipedia.org/wiki/SYSTRAN

SYSTRAN at that time was a rule-based machine translation and was comparable for european languages.

Where as Google's newer Translate and Microsoft Translator are statistical machine translators. (both trained with UN and EU documents that are available in several languages).

The future is probably a hybrid system (combination of rule-based and statistical) like the most recent SYSTRAN version.

I think Douglas Hofstadter has kept his idea of what AI is more constant than most folks.

And the domain of things we can empirically do better in the discrete task realm keeps diminishing ie. 'design', unsolvable capchas.

If we keep throwing morsels of cognitive capability to the machine, we're left wondering why the fuck we couldn't predict and develop Flappy Birds as a killer start-up.

This is where I find X-Men to be an awesome concept. What we are chasing in AI right now is Magneto (the might), but every time we reach there, we compare ourselves with Prof. X (the intellect).

I actually like this direction. Intellect is about predicting problems, and might is about solving them. It is better for humanity to stay on the intellectual side and let machines handle the hard work. Imagine a choice between AI propelled super efficient kill-bots vs. Skynet, if you will.

Reply on news.ycombinator.com