Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[counting]

Chatbot is an example of natural language processing and expert system combined to generate human-like response. It is a very old field in AI, date back to ELIZA in the late 60s. It's core principles haven't changed much since then, and is considered one of the easy implementation in AI (but not necessary good). It's indeed first created as a fancy way of showing how AI system could be in AI's early days (a toy or demo if yo like to call that). But it's limitation is quickly discovered. It took about 50 years to become like IBM watson. And it's still quite limited in many ways. (Expert system means it's only "expert" in certain area, and chatbot is only expert in replying chat through natural language phrasing rules, much like a therapist replying your words with your own words.)

P.S. Chinese room arguments has one serious problem that prevent it to become true. The combinations of possible words sequences constructing sentences and their responses are incredible large, it's known as exponential explosion. So large that even all the atoms in the universe can not hold this amount of information. Hence a chatbot with a quite limited database, will quickly reduced into "unnatural responses", simply because it runs out of phrases, and looping back into previous (stored) responses.

[alway]

Yep, in order to have an AI capable of intelligent conversation which isn't merely phrase salad you need at least some form of reasoning about texts' meanings, rather than the salad generated by things like Hidden Markov Models.

On another note, there is an interesting side effect of Cleverbot's training which apparently made itself known at the Loebner prize attempt. Apparently when they asked it about its own backstory, it had conflicting answers, as it pulled backstory from those who had previously spoken to it. That, really, is indicative of what Cleverbot is at it's core: little more than a strange mirror with memory.

On the topic of Watson, that AI is a bit more interesting. It actually does process language, search for meaning, learn and learn about its own learning. It learns which of its (400 or so iirc) algorithms work best in which situations and general categories of knowledge and weights them accordingly (learning about learning), it learns meanings of words through contextual clues based on what's around them, and then can use logic to discover relationships between words and information. It's not anything resembling strong/general AI, as it is a highly topical question-answer system, but it's a pretty damn good QA system.

[counting]

...
On the topic of Watson, that AI is a bit more interesting. It actually does process language, search for meaning, learn and learn about its own learning. It learns which of its (400 or so iirc) algorithms work best in which situations and general categories of knowledge and weights them accordingly (learning about learning), it learns meanings of words through contextual clues based on what's around them, and then can use logic to discover relationships between words and information. It's not anything resembling strong/general AI, as it is a highly topical question-answer system, but it's a pretty damn good QA system.

For someone who is interesting in machine learning, this is called "boosting" which is one kind of meta-learning. (Famous one like Adaboost, is a classifiers boosting which can be used in image recognitions).

And no matter how "natural" watson may look like, it's still just a very "specific expert" on database <-> natural language QA system. You can say it does NOT resemble the knowledge storage and learning system in a human brain. (Using only the principle like : ontology map)

And here it's official systematic structure lecture.  It even got a book : Final Jeopardy: Man vs. Machine and the Quest to Know Everything.

[Kay12]

Chatbot is an example of natural language processing and expert system combined to generate human-like response. It is a very old field in AI, date back to ELIZA in the late 60s. It's core principles haven't changed much since then, and is considered one of the easy implementation in AI (but not necessary good). It's indeed first created as a fancy way of showing how AI system could be in AI's early days (a toy or demo if yo like to call that). But it's limitation is quickly discovered. It took about 50 years to become like IBM watson. And it's still quite limited in many ways. (Expert system means it's only "expert" in certain area, and chatbot is only expert in replying chat through natural language phrasing rules, much like a therapist replying your words with your own words.)

P.S. Chinese room arguments has one serious problem that prevent it to become true. The combinations of possible words sequences constructing sentences and their responses are incredible large, it's known as exponential explosion. So large that even all the atoms in the universe can not hold this amount of information. Hence a chatbot with a quite limited database, will quickly reduced into "unnatural responses", simply because it runs out of phrases, and looping back into previous (stored) responses.

The idea of chatbots is actually much older, as they're basically what Alan Turing was studying in the 1950's - machines that provide humanlike responses.

And about the Chinese room argument - exponential explosion does not invalidate the Chinese room argument, but it does prevent it from becoming a practical consideration as building a comprehensive manual for humanlike conversation is impossible. However, the core of the Chinese room remains true - a machine that mimics human conversation perfectly but does it only by simple input/output mapping (checking the "manual") isn't any more intelligent than any other machine programmed to memorize stuff, despite the humanlike responses.

[counting]

The idea of chatbots is actually much older, as they're basically what Alan Turing was studying in the 1950's - machines that provide humanlike responses.

Theory wise. It's till late 60s, (digital) computers become really helpful tools in actual use and implemented the theory. A lot of the AI theory can all trace back to Turing, since he is the pioneer of AI after all.

And about the Chinese room argument - exponential explosion does not invalidate the Chinese room argument, but it does prevent it from becoming a practical consideration as building a comprehensive manual for humanlike conversation is impossible. However, the core of the Chinese room remains true - a machine that mimics human conversation perfectly but does it only by simple input/output mapping (checking the "manual") isn't any more intelligent than any other machine programmed to memorize stuff, despite the humanlike responses.

Chinese room argument is flawed (at least in its original form) and controversial in many ways. It's interesting to thought about, but not helping in understanding how AI actually works. Even hinder and put negative view in AI researches, and most who love the idea that strong AI can be implemented in our life time don't like this argument very much. (And against the romantic idea that an android can have "mind" or even "soul" as we human can. Or just damned even humans into mindless automatic machines.)

But for people who write AI related-programs (like myself using agent-based model) don't actually care it's philosophical meanings and implications. As far as I am concerned, if a thing works right the way it should be, it's intelligence enough. If an apple smells right and taste good, then there is no need to bother saying "artificial apple is not a real apple".

[Kay12]

So, would you explain what's wrong with the Chinese room argument? In my opinion it's a good way to highlight how a machine may appear to be intelligent without actually being so.

EDIT: So you decided to answer by editing your original post... just so you know, I like and study artificial intelligence and I believe machines can be truly intelligent - I brought the Chinese room argument up because it highlights one wrong approach towards it - building a "perfect" manual instead of making the machine actually reason about the proper response to each input.

[alway]

It's essentially a case of p-zombies. The problem with those being that, in a nutshell, it implicitly has the assumption of duality. Take, for example, the human mind. What p-zombies speculation does is assumes that the only observable pieces are the inputs in the form of sensory information and outputs in the form of behavior of some form. It thus boils down to circular reasoning: it starts with the assumption that experience is somehow different from the material reality; when in fact, every step, every grain of sand in the avalanche, is itself material and measurable. Every tiny fragment of intelligence is itself a part of the universe and can itself be measured. The falacy p-zombies falls into is falsely assuming we are somehow some sort of magical creatures which have experience and existence apart from the material world and which is somehow immaterial. Essentially that emergent or collective behavior is somehow more than the sum of its parts in a way which is more than mere metaphor; it's a classification error which stems from the way in which we interact with the world. It's similar to taking the temperature in a room. Temperature, as a thing in and of itself, doesn't exist. It's merely a measurement of the statistical distribution of the motion of particles. What p-zombies suggest is that one can have those same moving particles and somehow not have temperature.

[counting]

So, would you explain what's wrong with the Chinese room argument? In my opinion it's a good way to highlight how a machine may appear to be intelligent without actually being so.

EDIT: So you decided to answer by editing your original post... just so you know, I like and study artificial intelligence and I believe machines can be truly intelligent - I brought the Chinese room argument up because it highlights one wrong approach towards it - building a "perfect" manual instead of making the machine actually reason about the proper response to each input.

I'll make a post about why it is controversial in my opinion.

It has something to do with John Searle(He brought up this argument)'s strong disbelieve in machines / programs can ever be "constitutive of sufficient for minds to exist"(his own words and conclusions, He is quite a strong character). And it angers a lot of the strong AI believers. The original argument itself emphasize too much on the importance of understanding and "true minds", thus Chinese Room Arguments demonstrate that an AI will always be pure simulation, thus it's more or less like a "fake". (Using this intuitive argument, Searle is kind of calling a lot of researchers frauds, you can understand why it creates wide rages). In fact John Searle strongly oppose functionalism and computationalism, he thinks "the minds" can only exist in human messy chemical based brains (somehow mysteriously). Although his latter refined version, revised a lot of the original claims. (Sorry that I DO live through the age where the arguments is brought up, so I DO feel a little bit angry as well with Searle's arguments and responses, kind of brought up bad memories).

The strong AI community is in fact more against Searle himself, rather than the argument itself sometimes. In connectionist mind (like me, I studied Neural Network, and latter swam intelligence, leading to ABM), the way the Chinese Room works is probably not a proper analog to what a brain should work. The approaching or simply using raw computational power is just too "violent and inefficient" in my opinion. The way CPU actually works with very limited processing power in real life, may prove that unless we involve massive parallel computations, we may never achieve a functional AI that can behave naturally in human world.

P.S Searle will never claim himself as philosophical-zombie, and claim himself as real, but argue that Chinese Room is p-zombie.

[Kay12]

I understand why the Chinese room doesn't work as an argument against strong AI in general, but it certainly will work well against chatbots that actually try to build the Chinese manual... or would work if the possibilities of conversations wouldn't be too large to memorize anyway.

When it's about Chatbots attempting to mimic humanity, by the way, it actually makes a difference whether the bot uses actual reasoning or simple input/output mapping (with a few added mechanisms to prevent it from repeating itself too often). An example is the prime number stuff I mentioned earlier - a truly intelligent Chatbot would understand when being told about primes, or at least after seeing a few examples. Okay, primes are something that's fairly easy to hard code... but teaching concepts the creators of the AI hadn't thought of wouldn't work. Applies to practical concepts as well.

Personally, I'm very interested in neural networks. I'm not sure about them being especially practical, but sure as heck they're interesting...

[counting]

I understand why the Chinese room doesn't work as an argument against strong AI in general, but it certainly will work well against chatbots that actually try to build the Chinese manual... or would work if the possibilities of conversations wouldn't be too large to memorize anyway.

When it's about Chatbots attempting to mimic humanity, by the way, it actually makes a difference whether the bot uses actual reasoning or simple input/output mapping (with a few added mechanisms to prevent it from repeating itself too often). An example is the prime number stuff I mentioned earlier - a truly intelligent Chatbot would understand when being told about primes, or at least after seeing a few examples. Okay, primes are something that's fairly easy to hard code... but teaching concepts the creators of the AI hadn't thought of wouldn't work. Applies to practical concepts as well.

Personally, I'm very interested in neural networks. I'm not sure about them being especially practical, but sure as heck they're interesting...

NN is kind of falls into pure theoretical studies after the 90s. In practical application, its functionality is just as good (well actually as bad) as other means of classifiers. (The story of overly used all-powerful BPN). But NN is more than just simple classifier. (Data mining has much more efficient methods in clustering and classifying). Its power lies in with the learning ability and fault tolerant mechanism built-in into its structure. But a lot of performance comparison in papers only focus on how precious it can be. Which in most case not the best, and even kind of worse, and given it's extensive use of computation resources, makes it's impractical in most computational environment. (And giving it's unfair name)

But think about it that a NN's structure is independent with the question, and it requires no additional external variables. It truly is a black box like a human brain. And it's strongest power is never fully explored. But even in today's clusters of computers, running a NN program is still challenging. It's designed to run in massive parallel computations (as many as the links). So it contradicts with modern digital computer design philosophy. It's origin is analog computers, which runs with many parallel imprecise components. There are many aspects in NN that is inherent in its structure makes it an ideal rather than a strictly defined computer-science subset. Strictly NN is just it's math and models, but things like ontology models are actually the extension of connectionist ideal. I'd like to call NN one way of representing how human brain works in the language of computers science.

[Kay12]

By what I've understood, one of the best sides of neural networks is that they can learn to understand connections that are hard to express mathematically. However, the part about parallel computation is true, yeah. Neural networks are still interesting, though!


Funny story about neural network mirroring stupidity (or carelessness): A friend tried to teach a neural network to categorize photos - true if the photo contains a car and false if it doesn't. He spent several hours strolling around to get as many car pics as possible - varying colors, models, sizes, moving cars, stationary cars... as non-car control pics he used some of his old pictures - nature photos, camping trip pictures, meme pics, anything that had zero cars in it.

By what I've understood, it took him ages to teach the neural network, but eventually it did manage to classify all car pics correctly. The problem was that some non-car pics still returned true, despite not having anything a human could mistake for a car.

Eventually, my friend found out how his network recognized cars. He had had timestamping on when he was taking car pics while most of the other pics didn't have timestamps. The network thought that timestamp means a car. The timestamped non-car pics occurred rarely enough to make the network not lose the timestamp idea.

[counting]

By what I've understood, one of the best sides of neural networks is that they can learn to understand connections that are hard to express mathematically. However, the part about parallel computation is true, yeah. Neural networks are still interesting, though!


Funny story about neural network mirroring stupidity (or carelessness): A friend tried to teach a neural network to categorize photos - true if the photo contains a car and false if it doesn't. He spent several hours strolling around to get as many car pics as possible - varying colors, models, sizes, moving cars, stationary cars... as non-car control pics he used some of his old pictures - nature photos, camping trip pictures, meme pics, anything that had zero cars in it.

By what I've understood, it took him ages to teach the neural network, but eventually it did manage to classify all car pics correctly. The problem was that some non-car pics still returned true, despite not having anything a human could mistake for a car.

Eventually, my friend found out how his network recognized cars. He had had timestamping on when he was taking car pics while most of the other pics didn't have timestamps. The network thought that timestamp means a car. The timestamped non-car pics occurred rarely enough to make the network not lose the timestamp idea.

NN is good at non-linear analysis. Basically it just means that everything complex and hard to find patterns using linear equations. NN is based on patterns rather than symbols like binary representation. And the problem of over-fitting and false classified is a big problem considering "precision" in NN.

NN doesn't behave like any other deterministic algorithms (engineers love certainty and repeat performances), it behaves like GA (another famous time bioinspiration AI methods), you can never predict it could go (and go wrong) next. It's one giant black box. And this property often posses problems when trying to analysis it's performance. It's so "human" as it's so "stupid". And it's performance is also very "human". It varies greatly from data sets to data sets, even at different time (if it's a recursive structure). And it's initial state will greatly effect it's results. And it's famously for learning so slow as a human learning to play a piano. If you want to build human like AI, NN is probably a very good analog, but from engineering point of view though, it's full of flaws.

[Kay12]

Their lack of practical uses (as of now) doesn't make neural networks any less interesting

[DrPoo]

Their lack of practical uses (as of now) doesn't make neural networks any less interesting

Yes there is, if we want some kind of uniform servant drone thing, a neural network would be an easy way to teach it when to fire off its servo's and when to beep, when to activate that periphal, etc.
As far is i know, lawnmover bots uses some kind of NN. I am not quite sure..
But yeah, somewhat cool at best, useless at worst.

[Kay12]

At least for now... I can only imagine how cool neural networks may get once someone creates a computer that can run them efficiently.

On the other hand, for most purposes, you'll get along just fine without them.