Wednesday, 12 August 2015

Artificial Intelligence (AI): All About Algorithms?

...Thoughts-Computations-Rules-Algorithms...

Many "cognitivists" believe that the brain is a computer. (Sometimes they say “a kind of computer”.) Thus, as a result of this belief, they attempt to discover the computational processes which enable such things as perception and learning. However, expressed in that manner (as it often is), things are a little unclear.

Such cognitivists believe that the brain is also a machine – a computing machine.

Computationalists (computationalism is a branch of cognitivism) claim that all thought is computation. But what does that mean? Are the words 'thought' and 'computation' are virtual (or literal) synonyms?

This is more clearly the case because it seems that almost all conscious processes in the brain are deemed to be thoughts; and thus also deemed to be computations.

That not only includes the thought that 1 plus 1 equals 4 or that Snow is white; but also the rotation of a mental image in the mind, imagining the smell of a rose and so on. Then again, if rotating a mental image is classed as a thought, then why can't it also be classed as a computation? Especially since, in computationalism, they appear to by synonyms.

It all now depends on what we mean by the word 'computation'.

For a start, we can make the following claims:

i) All of a computer's processes are computations.

ii) Not all conscious human mental processes are computations.

Similarly, we can say:

iii) Many human mental processes are thoughts.

iv) No computer computations are thoughts (i.e. because thoughts have semantic content, intentionality, reference, etc.).

Rules Rule, Okay?

Is everything that happens consciously in the mind a computation? Or, perhaps more tellingly, is it all rule-governed? Jerry Fodor doesn't think so. He says that

“some of the most striking things that people do – ‘creative’ things like writing poems, discovering laws, or, generally, having good ideas – don’t feel like species of rule-governed processes”.

The way Fodor puts his position doesn't really help matters Sure, such things may not “feel like species of rule-governed processes”. However, that doesn't mean that they aren't rule-governed processes. Far from it. This is the same phenomenological approach that's applied to free will. Here again most people “feel like” they have free will. Though, on close inspection, that claim (about what things feel like) amounts to almost nothing.

On Fodor's behalf it can now be asked what something's being rule-governed could possibly mean in the varied contexts of “writing poems, discovering laws, or, generally, having good ideas”. Are these disparate things really united by the following of rules (if at the non-conscious level)? Well that would depend on what's meant by the words “rules-governed”.

We can take this somewhat further.

If writing poems, discovering laws and having good ideas are rule-governed, then these creative processes must be following some kinds of algorithm. And following on from that, they must be computable. What's more, this could mean that these processes are rote in some (or sometimes all) respects. Not in the sense of conscious acts of rote learning (or memory); but in the respect that the brain (or physiological system) has 'acquired' certain modules/faculties/etc. - or that such things are innate.

Haven't we simply moved from one technical term (i.e., 'rule-governed') to two more – 'algorithms' and 'computable'? After all, the words 'algorithms' and 'computations' can both be be cashed out in terms of following rules or being rule-governed.

So let's quote a definition of the word 'algorithm' as it's specifically used in reference to computers:

“An algorithm is basically an instance of logic written in software by software developers to be effective for the intended 'target' computer(s) to produce output from given input (perhaps null).”

The mention of 'logic' (along with the very mechanical way of describing both what an algorithm does and how it comes to be) seems to make Fodor's earlier claim a little more convincing. Can we say that “an instance of logic” (or instances of logic) is required to “write poems, discover laws, or, generally, having good ideas”? Yes, we can! It's certainly the case that - in a limited sense - instances of logic/algorithms will be involved in these processes. The thing is, it surely can't be said that it's all about logic or algorithms.

We can now say that logic or algorithms can be applied to some things (or to all things!) which aren't themselves logical or algorithmic.

Bad Computers

Kurt Gödel is often brought into the picture in order to show us what humans have and what computers (ostensibly) don't have.

For example, there's much talk about human brains having a "rule-free flexibility" and “unlimited mathematical abilities”. Thus there's also talk about “intuition” and “direct insight”. Of course these abilities can be seen to run free (conceptually speaking) of other things that computers don't have: such as qualia, emotions and suchlike. Then again, some would say that they all form a Gödelian package.

In concrete terms, there's the argument that humans can solve computational problems which computers can't solve. (Note that this hasn't got anything directly to do with computers not being able to write poems or have an orgasm.) This Gödelian claim that “no such limits apply to the human intellect” is, as Alan Turing argued in 1950, often “merely stated, without any sort of proof”.

In any case, various consequences are put forward as being a result of computers not having our (as it were) Gödel faculty. They include the fact that most computers crash for trivial reasons (e.g., because of faulty software or bad input). This is said to be due to the rule-fixated nature of computers; unlike human beings who have (as stated) a Gödel faculty.

All this is seen to be a direct result of computers needing a rule or algorithm for literally everything they do. More concretely, computers show no intuition or insight; and, in most cases, they don't learn from their mistakes or learn not to make mistakes. (Though this isn't true of all computers or even all aspects of each computer.)

Here again philosophers stress human uniqueness. Hubert Dreyfus, for example, argues that there are many examples of mental activity and behaviour that aren't a question of following rules. As Dreyfus himself puts it, computers lack the “immediate intuitive situational response that is characteristic of [human] expertise”. Consequently, persons

“must depend almost entirely on intuition and hardly at all on analysis and comparison of alternatives”.

Basically, some people argue that these Gödelian things can't be programmed into a computer.

The science writer John Horgan also tells us what computers are bad at. He writes:

“ Computers may excel at precisely defined tasks such as mathematics and chess.... but they still perform abysmally when confronted with the kind of problems – recognising a face or voice or walking down a crowded pavement – that human solve effortlessly.” (1996).

To state the obvious, the above are all programming problems. And they're programming problems because the number of variables the computer (as well as a person) needs to take into account when it comes to “recognising a face or voice or walking down a crowded pavement” are huge (or indefinite) in number. However, persons, it can be argued, don't (really?) need to be programmed in these cases: they react situationally. That is, persons can act upon - and react to - novel situations; even though (it can be said) these situations aren't entirely novel.

The philosopher George Rey also states the case that computers don't have a full logical package. He writes:

“Intelligence requires doing well under non-ideal conditions as well... But performing well under varied conditions is precisely what we know existing computers tend not to do. Decreasingly ideal cases require increasingly clever inferences to the best explanation in order for judgements to come out true; and characterising such inferences is one of the central problems confronting artificial intelligence...” (1986)

Here again we see that in all the cases in which a computer doesn't have a rule or algorithm to follow, then it doesn't know what to do. Of course you can create rules which tell a computer what to do when there are no existing rules; though that would depend on the nature of these meta-rules as well as upon the new conditions the computer is facing.

To sum up in the language of logic: computers aren't very good at “inferences to the best explanation” when they find themselves in “non-ideal conditions”.

Good Computers

Nonetheless, all sorts of new factors have been added to computers to simulate intuition or Gödelian intelligence. Such things as quantum computers based on “entangled qubits”, the introduction of random factors (e.g., annealing approaches) and hardware neural nets have been added into the computer-pot.

In any case, we already know about computer randomness. Even von Neumann machines can modify their own programmes (i.e., they can learn). That means that some of their responses (or output) are unpredictable. All this is achieved, in general, by equipping a computer with certain random elements which the computer can work on to produce outputs which are unexpected (i.e., which have moved beyond the programmed data). Indeed all this was theorised about by Turing as long ago as 1938.

Even with early Turing machines there was a requirement that such machines be able to follow their own rules or show what some people (at the time) called “initiative”. It was said that a programmer could engineer an element of randomness into the computer (or into the programme). That was what Alan Turing himself tried to do with his “Manchester computer” (1948-50). That meant that such randomness (as it were) would bring about “intuition” (or initiative) in the Turing machine – or even free will!

So when (not if):

i) A random element is introduced into a Turing machine (or a computer),

ii) and that computer manages to follow rules not laid down by the programmer,

iii) and as a result of that it solves its own problems,

iv) then that computer has learned something of its own accord or it even has “intuition”.

Thus there's no “appearance” about it! In this limited respect, the computer is free from its programmer. Or it has a “will” which is independent of its programmers. This isn't to say that it has either a mind or a (free) will in the human sense; though the independence (or freedom) is certainly real.

I think it would also be correct to say that a Turing machine “could have done something else” with the same input. That is, the same random change (mentioned by Turing) to the Turning machine can have different results in terms of what it produces. (E.g., a different calculation or even a different action – though a calculation is an action of sorts.)

More specifically in terms of today's computer programmes, there's what is called “machine learning” in which computer programmes have the ability to “self-modify”. These include programmes which themselves include ensemble learning, current-best-hypothesis learning, explanation-based learning, decision-tree learning, reinforcement learning, Bayesian statistical learning, instance-based learning and so on.

Despite all that, it's still said (by some) that none of these things (not even collectively) produce a Gödelian mind. That's because all these additions can still be reduced to Turing machines (along with their limitations). Sure, they make computers much better; though it's still said that they don't make them Gödelian.

References

Dreyfus, Herbert. (1992) What Computers Still Can't Do