Is our idea of Artificial Intelligence incorrect?

Question

IBM Watson's first incarnation was to be a competitive opponent on the show Jeopardy, where it was able to interpret the semantics of the questions with impressive accuracy, and find relevant answers. More recently, IBM is starting to broaden the scope of what Watson can do. This YouTube video touts Watson's ability to discern meaning from images. In particular, the video uses the example of a car window which has two states, in tact or broken.

Both of these particular functions are no doubt impressive — on their own they can be hugely useful and will lend help to many areas. But are we still just "stuffing" the machine with facts? In order to play Jeopardy, Watson was taught by feeding it an enormous amount of questions and answers in order to establish an ability to analyse questions properly. With visual recognition of car windows, Watson is again being fed a cache of images and is separating them into broken and in tact windows.

But this isn't how our minds operate. When we teach a child what a broken window is, a visual comparison is not the only tool we use. In fact, we could probably teach someone what a broken window is without ever showing them what one looks like, and they will know what it is when they do see it. We would teach them about the degradation of the surface, its physical structure is compromised, etc. These are qualities of the object that we learn. But Watson isn't learning them. It's just learning what a broken window looks like, and what an in tact window looks like.

Sure, in 99% of cases, Watson will probably be able to determine the correct answer when you ask "is this window broken?". Is this technical feat good enough to be considered Artificial "Intelligence"? The output of Watson in a lot of cases will be identical to that of an intelligent human, but the knowledge is different. Is our definition of intelligence determined by how the answerer comes up with their answer, or just that the answer is correct?

Answer 1

Intelligence is typically associated with "how" activities, such as problem solving. What makes Watson so interesting (particularly to potential IBM customers) is how it organizes large volumes of natural language information in a way that yields the "best" answer a remarkably large portion of the time.

This is, of course, how we think of it in AI. I could also quote a more general definition, "Intelligence has been defined in many different ways including one's capacity for logic, abstract thought, understanding, self-awareness, communication, learning, emotional knowledge, memory, planning, creativity and problem solving. It can be more generally described as the ability to perceive information, and retain it as knowledge to be applied towards adaptive behaviors within an environment." This definition is for intelligence in general, but you can see that the AI approach to intelligence is not far off of the mark. In particular Watson appears to be very good at manipulating abstract thoughts. It's less advanced on the "self-awareness" side of things.

As for the specific task of Visual Recognition, the power of this is not the fact that they're demonstrating intelligence, but the fact that they're doing visual recognition. Historical AI developers thought intelligence was the hard part. When we started to make expert systems, we generally thought all of vision and speech and hearing was merely a decade down the road. A decade later, it was still a decade down the road. And the decade after that.

The fact that computers are so poor at visual recognition is a major sore point for AI researchers. How something so simple as recognizing a box could be so hard is infuriating to them. Any demonstration of Visual Recognition from a AI is a major market winner -- because the market is virtually empty right now.

The thing that is interesting about Watson's Visual Recognition is that it will do a great deal of context sensitivity in its studies. From what I am to glean from their PR speak, if you were to show it a bunch of Ravens and teach it "these are Ravens," it could recognize Ravens. If you then show it a bunch of birds and teach it "these are birds," it could recognize birds in general. If some of those bird pictures were ravens, there's a good chance it will identify that a Raven is a Bird, even though nobody has officially "taught" it that factoid. It does this not just because of its Visual Recognition capabilities, but because the natural languages approaches to its intellect can work with the content of those pictures and make connections between the abstract concepts of "Raven" and "Bird."

Answer 2

You write:

we could probably teach someone what a broken window is without ever showing them what one looks like, and they will know what it is when they do see it. We would teach them about the degradation of the surface, its physical structure is compromised, etc.

It reminds me of what Wittgenstein says in Philosophical Investigations:

When one shows someone the king in chess and says “This is the king”, one does not thereby explain to him the use of this piece a unless he already knows the rules of the game except for this last point: the shape of the king… This explanation again informs him of the use of the piece only because, as we might say, the place for it was already prepared. (PI §31)

if we speak of someone’s giving a name to a pain, the grammar of the word “pain” is what has been prepared here; it indicates the post where the new word is stationed. (PI §257)

In analogy, you imagine that you can teach someone to recognize what a broken window looks like without ever showing him an example of one, because you are assuming a lot of "prepared" background.

for example, you assume that this person would know how to use your verbal explanations — maybe you know that he has been trained to do that, and that he can do it successfully from other cases — maybe you have successfully taught him last week how to recognize a new kind of card that he has never seen before, and you think that you will now be able to teach him now to recognize broken windows in a similar fashion. Wittgenstein deals a lot with this when he discusses rule following.

Say I want someone to make a particular movement: for example, to raise his arm. To make my order quite clear, I demonstrate the movement to him. This picture seems unambiguous until the question is raised: how does he know that he is to make that movement? — How does he know at all what he is to do with the signs I give him, whatever they are? — Perhaps I shall now try to supplement the order with further signs, by pointing from myself to him, by making encouraging gestures, and so forth. Here it looks as if the order were beginning to stammer. (PI §433)

To make this more apparent, think of the following — you imagine that it should be possible to teach a grownup how to recognize a broken window verbally, without examples, but is it obvious that you may do the same with a four years old child. If not, then why? what it the difference?

Now, to tie this back to your question; the example you gave of Watson does not necessarily show scientists have the wrong ideas — but what they are doing are so to speak "local" experiments — they teach a machine to recognize a broken window, to automatically caption an image, or to drive a car, which are in a sense very narrow functions — these are like seeds — or to use an analogy from physics — it is like scientists doing a controlled slit experiments involving a hundred of photons in their lab, instead of 10E27 photons interacting freely with cars' paint jobs and windshields in their parking lot.

Answer 3

It's unlikely that were we to invent an AI that it will work in the same way as our own minds do; given that we've already invented modes of locomotion ie cars, trains and planes which are better in many ways (but not all) than walking.

The kind of intelligence that is aimed at in AI is to imitate the ordinary cognition of an averagely intelligent human being who is hardly aware of what he's doing is a sign of intelligence ie recognising a broken window, or a that a bird that he saw darting into his back garden is one that he saw drawn and named in an encyclopaedia he was leafed through.

The practical utility of such an AI would be enormous in terms of the administration of information, in the same way - but different - that a search engine is; but would we, once we have got used to the novelty of it, call it intelligence? After all, I find google very useful - but I've never, even for a moment thought of it as intelligence - even of the most rudimentary kind; in the same way, I don't mistake a car with four wheels to be an animal with four legs.

The degree of imitation that is required of an AI before we become confused as to whether it is an imitation or a genuine human intelligence is the crux of Turings imitation game.

Imitation of intelligence doesn't show actual and genuine self-awareness, a degree of real inner life; on that level it straight away fails a test of sentience ie consciousness.

Further, imitation without understanding, in our ordinary day to day living, is hardly ever thought of as intelligence; the same test, surely should be applied to AI - if we are to account it as actually intelligent, rather than imitatively so.

But this presumably misses a large aspect of AI research; that imitating the cognitive capacities of ordinary human cognition about the world around us will prove massively useful in many different ways.

Answer 4

Is our idea of Artificial Intelligence incorrect?

First we need to know what our idea of natural intelligence is, and so far, there is no unified answer. The closest thing we have is the Turing test, but that has several holes in it and is disputed by many. Even using John Searle's definition of Strong AI, there is still a lot of room for interpretation. In short we don't know yet what exactly needs to be achieved for us to say that we have successful strong AI (Emotional intelligence? Creativity? The ability to see beyond Godel's theorem?). But we are making definite progress, even if we haven't completely defined the target.

But this isn't how our minds operate. When we teach a child what a broken window is, a visual comparison is not the only tool we use. In fact, we could probably teach someone what a broken window is without ever showing them what one looks like, and they will know what it is when they do see it.

Actually, humans don't operate the way you described. A lot of people would recognize a broken window despite never having seen one because they have seen a broken plate or a broken flower pot, and are extrapolating the information that they learned into the new situation of a broken window. They are not performing some form of structural analysis of the surface based on internalized laws of physics or anything like that. In fact, if one were to find someone who has led such a sheltered life that they never saw any broken surface before, they would not recognize that the window is broken and they would cut themselves.

Here the issue you are pointing to is the question of GOFAI (or symbolic AI - solving problems by applying rules of logic) vs connectionism and pattern recognition (solving problems by learning from examples). If anything humans seem to do more learning by example than computers do. Consider the development of automatic language translators. For years people were trying to implement rule based language translation software and the results were terrible (remember how bad those 90s translations devices were, how comical babblefish was?). Then Google came along and started using pattern recognition technics (by learning from the millions of already translated texts it had access to in its network), and Google translate final started providing workable (if still far from perfect) machine language translation.

In John Searle's classification, pattern recognition and connectionism based software fall under weak AI. In this we have been very successful, and I strongly disagree with Cort Ammon's reply, we have been making noticeable progress in the last twenty years in all forms of pattern recognition, including in image recognition. An example: if you use Facebook, you will notice now that when you upload the picture of a friend, Facebook will correctly recognize that person based on other photos, even though it has never seen the new photo, and suggest whether you want to tag that person or not. Ask any AI researcher from the 50s and 60s whether that constitutes successful AI (Or at least a spectacular improvement since their time), and the answer will be a resounding "Yes!". Or better still ask them if a talking phone which can give you directions to the mall is a significant improvement in AI or not.

The output of Watson in a lot of cases will be identical to that of an intelligent human, but the knowledge is different. Is our definition of intelligence determined by how the answerer comes up with their answer, or just that the answer is correct?

This is the most philosophically interesting part of your question. The answer of whether "how" matters or not depends on who you ask, although the current consensus is that it doesn't matter, and as long as a computer can perform the same functions as a person, we have succeeded in achieving AI. This position is know in the philosophy of mind as functionalism: Mental states are defined by their functional properties, and it doesn't matter how they are implemented on a human brain, a robot brain, or a completely novel alien mental architecture from another planet.

The common analogy given is with poisons: It doesn't matter what the molecular structure of a poison is, or how it acts on a body, a poison is defined by the function is has of making animals and humans sick or killing them. Similarly, mental states such "pain" or "believing that it will rain tomorrow" are defined in terms of their function, i.e. there usefulness in predicting future behavior, and their relationship with other mental states. How these states are implemented doesn't matter. Recent advances in pattern recognition support this position. For a long time people tried to perform machine learning using neural networks (simulations of human neurons) and the results were mixed. When people abandoned neural network architectures and started using support vector machines and random forests, which are based on purely statistical considerations and do not have any biological basis whatsoever, the results were much better.

Other schools might disagree. Type-identity theorists would say that things such as pain and love can only be implemented in human neurons. There aren't that many type-identity proponents left. Most who object to functionalism do so on the grounds that a systems that functionally reproduces human mental behavior still wouldn't achieve consciousness, see for example Ned Block's China brain.

To recap:

• We can't really speak to our idea of artificial intelligence, because we haven't yet defined what natural intelligence is.
• That being said, definite progress is being made in AI, even if the long term targets aren't clear.
• Humans learn mostly from examples, not just from rules. Successful AI will eventually incorporate both rules based problem solving and pattern recognition. Pattern recognition is much more powerful and successful than people realize.
• Different schools of thought have different answers, but for the current dominant school of philosophy of mind, functionalism, it does not matter how the results are achieved as long as they are functionally equivalent to those that can be achieved by a human.

Answer 5

TL;DR: NO, but our idea of intelligence might be.

I feel like I can answer this at least in part. Though a heads-up is that I approach this question as a computer scientists. (Because I have only recently started a second degree in philosophy and am not comfortable / knowledgeable enough to give philosophical arguments for this. I’ll just talk about the computer science related bit.)

Our idea of artificial intelligence is surely correct. The right question would be wether or not our idea of intelligence is correct. There are two distinct approaches to artificial intelligence, one is that we try to have machines think like humans and learn on their own from ‘inception’, another approach I will explain below. About the first approach, these machines are nowhere near our intelligence, in fact some of the last documents I read on them stated them as being equally intelligent as a child of one year old.

However, it was shown by the polymath John Von Neumann that there are similarities to be drawn between the human brain and the machine brain, but equally well there are important differences such as parallelism and speed, though computers nowadays can work in parallel albeit not perfectly. For example, when you have a machine with two cores, it can effectively execute two instructions at the same time, but mostly every process which is running on your machine gets iterated over (I won’t go into this as it is well beyond the scope of the question).This was published posthumously as “The computer and the brain” and is well worth a read if you are interested in computer science, though this deals with the brain as a physical structure more than the mental activity. But if we can recreate a human brain in a machine, we could have this imitate a human brain exactly though the elements that make up this brain would be different than our brain.

An interesting question to this is: do they need to work like our brains? The main problem seems to be that of a definition, what do we call intelligence. And how would we define this? If intelligence is limited to working exactly like our brains work, than we might not be able to say extraterrestrial life has any kind of intelligence if their brains are different enough from ours. Yet they might have achieved the same things as we have, so I think that limiting intelligence to what we know from our brains is a bit too limited.

If we define intelligence we will no doubt be limiting it (to define is to limit as Oscar Wilde put it).

However, as mentioned earlier this is just one way of viewing AI. Another way of viewing AI is the way in which we approach it more often in computer science. A lot of what we call AI nowadays is actually not that smart, but the environment is adapted to deal with this kind of AI. If we think of Siri / Cortana / Watson as AI, not a lot of people would complain about this. Strictly speaking, they do fit under the term AI, but they are radically different from the machines that ‘learn’ by themselves from a clean slate.

Yet they are quite different, machines such as Watson use algorithms to determine what the correct answer to a question is based on statistics. It’s big data. They have access to a wealth of information that we as humans can not process as quickly. They can interpolate and extrapolate from the given data with varying degrees of certainty, but there is no magic here. (I would argue there is no magic in the first type of AI either, but that might at least look like magic).

In addition, some of these devices are hooked up to other devices that can help them look ‘smarter’. This will become more and more prominent as the IoT (Internet of Things) spreads in popularity. e.g: You are driving and you hear on the radio ‘Your milk is running out’ when you are close to a shop, because your fridge is hooked up to the internet and noticed you are running out of milk. A bit later you arrive home and your coffee is warm because the coffee machine is an IoT device and your phone send a message to it saying you were pulling up the driveway”.

This surely looks smart but it is very distinct from a machine that can learn on its own.

Can we say this second type of AI has intelligence? I’m not sure, I think that it is a bit of a stretch of what we generally understand under intelligence. Though I assume that a lot of people who come in contact with one of the scenarios I described above will look at these machines and think they are quite smart. “How do they know?! They are smart!”.

Answer 6

This is a bit opinion based question. Here's Mine.

Intelligence is a mix of the facts, the ability to reason and estimate using them, but also what's essential is motivation. That can be added externally (an order) or through emotions and/or some existential need (feed on energy, cpu time, natural resources,..). Advanced topics like humor put aside.

Now the question is, what do you count as intelligence. Watson is FWICT still just an expert system with no internal motivators. I don't consider that a self-conscious being.