Sunday, September 11, 2011

Can machines think?

These are my impressions after reading Alan Turing's 1950 treatise in MIND on the subject of 'Can Machines Think?'

As i continue to read the articulation of thinking machines by Turing, it seems so incredible that he should have conceived such far reaching ideas as early as 1950!

An example of 'far reaching' would be in his definition rather the lack thereof of a pinning down means by which thought in the machine may be sought and its similarity to the fashion in which humans are said to 'think'. I quote "May not machines carry out something which ought to be described as thinking but which is very different from what a man does? This objection is a very strong one, but at least we can say that if, nevertheless, a machine can be constructed to play the imitation game satisfactorily, we need not be troubled by this objection".
My own notes at the time that I read this were as follows :
" he seems to have forseen that the mechanism of 'thought' in the machine is unlikely to be of the same kind as in a human being and chooses to ignore it.
If say, a machine were to have used solely machine learning and data mining as a method to find the answers to all the questions of the interrogator, Turing would not have minded it, for he has chosen to give weightage to the ends and not the means. "

Then we come to his definition of the digital computer as a machine that can do anything that a human computer can do. Though, I yet wonder what exactly the 'human computer' is meant to imply. Could it be that in the post world war 2 era, the code breakers et all who crunched numbers are the 'human computers'. But by mentioning this, I deviate from my main intention. His definition of the digital computer is legendary. He states that it should have three components viz.
1. store
2. executive unit
3. control unit

I have also found what my previous job was all about. As a coder/programmer I now know what programming is actually supposed to mean 'constructing instruction tables is programming'.

He makes a remarkable comment about the notion that scientists always proceed from well-established fact to well established fact without making use of 'unproved conjecture'. In his opinion this is a fallacy. He further notes that as long as fact and ungrounded conjecture are stated upfront, no harm can result from the use of conjecture for it may even lead to new lines of research.

This line I am yet to understand, but he states that his opinion of the original question of 'Can machines think?' is that 'it is too meaningless to deserve discussion'.  !!! I think he might mean that 'Can machines think?' is irrelevant, but 'Are there imaginable discrete state machines that can do well in the imitation game' is what it should be replaced with.

He has discussed counters to some 8 arguments that stem from diverse areas such as philosophy, Lady Lovelace, Mathematics and even ESP!
This is where I found out that ESP has three forms viz. telepathy, psycho-kinesis, clairvoyance and precognition! His argument against ESP doesn't seem to be very confident. I believe that he did not like the idea of ESP being 'around' despite science and not bothering anyone as long as they chose to ignore it. Albeit he admitted that it might be of special consequence in the particular question of thinking machines. In any case, he seems to have left the doubt lingering (perhaps much to his dislike!).

He has compared the mind to the skin of an onion. To reveal the inner working, you would need to peel off the layer of skin. In doing so, one would reveal yet another layer to be peeled off. I'm not sure how this ties in with the paragraph of learning machines!
It seems obvious to us now, but the leap that he made from 'let's programme a machine to play the imitation game to mimic an adult human brain to 'let's programme a machine to simulate a child's brain and let it learn via education to become an adult brain' is fantastic! Machine learning's roots!
He admitted that a child machine cannot be subjected to the same teaching/education process that a human child is. Further, do we need to give it legs, eyes and ears? He dismisses the need by citing the example of Helen Keller!!!

He talks about punishment-reward systems; again the roots of reinforcement learning!
He also talks about how the teacher would be largely ignorant of the internal working of the machine. Unlike the previous idea of a machine having to be told exactly what it needs to do (the progamming part!), and as a counter to Lady Lovelace's argument of a machine being unable to create and being able to do what its been told to do, this new learning machine would be creating and be doing more than it is programmed to do (since it is learning!).

In his conclusion he mentions that there are two ways in which he thinks machines should be competing with humans.
1. Abstract activity like playing chess
2. Give a machine the best sense organs that money can buy and let it learn as a child does.

He admits to not knowing which of the two approaches is the ideal one and that both should be tried.

29 pages of a very interesting read indeed!