I’ll second the recommendation for Tom Mitchell’s book (although it has been a long time since I have read it and I have moved away from the machine learning philosophy since).
Are you going to go on to mention that the search in a finite concept space can be seen as a search the space of regular languages and therefore a search in the space of FSM? And then move onto Turing Machines and the different concepts they can represent e.g. the set of strings that have exactly the same number of 0s and 1s in.
Hmm, lets fast forward to where I think the disagreement might lie in our philosophies.
Let us say I have painstakingly come up with a Turing machine that represents a concept, e.g. the even 0s and 1s I mentioned above. We shall call this the evenstring concept, I want to give this concept to a machine as it I have found it useful for something.
Now I could try and teach this to a machine by giving it a series of of positive and negative examples
It would take infinite bits to fully determine this concept, agreed? You might get there early if you have a nice short way of describing evenstring in the AIs space of TM.
Instead if we had an agreed ordering of Turing machines I could communicate the bits of the Turing machine corresponding to evenstring first and ignore the evidence about evenstring entirely, instead we are looking at evidence for what is the TM of! That is I am no longer doing traditional induction. It would only take n bits of evidence to nail down the Turing Machine, where n is the length of the turing machine I am trying to communicate to the AI. I could communicate partial bits and it could try and figure out the rest, if I specified the length or a bound on it.
If you want to add evidence back into the picture, I could communicate the evenstring concept initially to the AI and make it increase the prior of the evenstring concept in some fashion. Then it collect evidence in the normal way, in case we were wrong when we communicated evenstring in some fashion, or it had a different coding for TMs.
However this is still not enough for me. The concepts that this could deal with would only be to do with the outside world, that is the communicated TM would be a mapping from external senses to thing space. I’m interested in concepts that map the space of turing machines to another space of turing machines (“‘et’ is the french word for ‘and’”), and other weird and wonderful concepts.
I’ll leave it at that for now.
Actually I’ll skip to the end. In order to be able to represent all the possible concepts (e.g. concepts about concept formation, concepts about languages) I would like to be able to represent I need a general purpose, stored program computer. A lot like a PC, but slightly different.
I’ll second the recommendation for Tom Mitchell’s book (although it has been a long time since I have read it and I have moved away from the machine learning philosophy since).
Are you going to go on to mention that the search in a finite concept space can be seen as a search the space of regular languages and therefore a search in the space of FSM? And then move onto Turing Machines and the different concepts they can represent e.g. the set of strings that have exactly the same number of 0s and 1s in.
Hmm, lets fast forward to where I think the disagreement might lie in our philosophies.
Let us say I have painstakingly come up with a Turing machine that represents a concept, e.g. the even 0s and 1s I mentioned above. We shall call this the evenstring concept, I want to give this concept to a machine as it I have found it useful for something.
Now I could try and teach this to a machine by giving it a series of of positive and negative examples
0011 +, 0101 +, 000000111111 +, 1 -, 0001 - etc...
It would take infinite bits to fully determine this concept, agreed? You might get there early if you have a nice short way of describing evenstring in the AIs space of TM.
Instead if we had an agreed ordering of Turing machines I could communicate the bits of the Turing machine corresponding to evenstring first and ignore the evidence about evenstring entirely, instead we are looking at evidence for what is the TM of! That is I am no longer doing traditional induction. It would only take n bits of evidence to nail down the Turing Machine, where n is the length of the turing machine I am trying to communicate to the AI. I could communicate partial bits and it could try and figure out the rest, if I specified the length or a bound on it.
If you want to add evidence back into the picture, I could communicate the evenstring concept initially to the AI and make it increase the prior of the evenstring concept in some fashion. Then it collect evidence in the normal way, in case we were wrong when we communicated evenstring in some fashion, or it had a different coding for TMs.
However this is still not enough for me. The concepts that this could deal with would only be to do with the outside world, that is the communicated TM would be a mapping from external senses to thing space. I’m interested in concepts that map the space of turing machines to another space of turing machines (“‘et’ is the french word for ‘and’”), and other weird and wonderful concepts.
I’ll leave it at that for now.
Actually I’ll skip to the end. In order to be able to represent all the possible concepts (e.g. concepts about concept formation, concepts about languages) I would like to be able to represent I need a general purpose, stored program computer. A lot like a PC, but slightly different.