3 ms·
In the interest of pruning this conversation a bit I will not continue the discussion about GPT-3. Apologies, but this thread is growing too fast and I don't ha
by YeGoblynQueenne 6y ago
In the interest of pruning this conversation a bit I will not continue the
discussion about GPT-3. Apologies, but this thread is growing too fast and I
don't have the time to give your comments the attention they deserve. I am
happy for you to have the last word in that matter.
>> These techniques were invented from the field of AI, but that does not mean
they remain in the field of AI.
Like I say above, it is pretty uncontroversial that these approaches are part
of the field of AI research. You can consult wikipedia or e.g. the calls for
papers from major AI conferences, AAAI and IJCAI, if in doubt.
So I have to ask again, why do you say these approaches are are not in the
field of AI research? According to whom? And based on what?
I would please like an answer to the above question.
Further, I can certainly point you to successes of symbolic AI, which you say
don't exist. For one thing, the entire fields of automated theorem proving,
planning, search, game playing, knowledge representation and reasoning, etc.
that you say are "not AI", but are like I say still active and still state of
the art in their respective tasks. These are certainly successful- they have
produced systems and techniques that still work best than any alternative and
actually quite well.
For examples of specific systems that were successful in their time, see Logic
Theorist [1] that proved 38 of the first 52 theorems in Principia Mathematica;
Chinook [2], the first computer program to win a world championship against
humans (in checkers/draughts); Deep Blue [3], the first AI system to defeat a
human grandmaster (Garry Kasparov) in chess; MYCIN [4] the first AI system to
outperform human experts in disease diagnosis (specifically, diagnosis of
infections); and so on.
Of course these systems have been superseded - but they were successes
nonetheless. Another reason to learn the history of AI is to become aware of
those systems- they, indeed, were "there".
Again I have to ask you- where does your knowledge of AI come from? When
you make such strong statements about what works and what doesn't, what failed
and what succeeded, are you sure you are well informed? Do you draw your
knowledge from primary sources, or are you trusting the opinions of others who
claim to be experts- but may not be (like in the article above)?
>> How would you tackle this with ILP?
Below I've defined the problem in the format expected by Louise [5]:
?- list_mil_problem(ordered/3).
Positive examples
-----------------
ordered([a],[b,c],[d,e,f]).
Negative examples
-----------------
[]
Background knowledge
--------------------
shorter/2:
shorter(A,B):-length(A,C),length(B,D),C<D.
Metarules
---------
triadic_chain metarule 'P(x,z,y):- Q(x,z), R(z,y)'.
true.
Given this problem definition, Louise can learn the following (Prolog) program:
?- learn(ordered/3).
ordered(A,B,C):-shorter(A,B),shorter(B,C).
true.
To explain, shorter/2 is a predicate defined as background knowledge by me.
triadic_chain is a metarule, a second-order clause that provides inductive
bias. length/2 is an ISO Prolog predicate.
Like I say, this is a trivial problem, not least because its solution is easy
to figure out and the background knowledge and metarules are trivial to define
by hand. Louise can also perform predicate invention to define new
background knowledge (kind of like inventing new features) and also new
metarules. That is to say, Louise can learn the shorter/2 and length/2
programs, also from very few examples- and then reuse them as background
knowledge. But showing how to do that would make for a larger example. I'm
happy to oblige if you are curious.
I should point out that there exists no neural net approach that can learn the
same (or a similar) program from a single positive example- not least because
neural nets cannot make use of background knowledge (i.e. a library of
programs from which to build other programs).
__________________
[1] https://en.wikipedia.org/wiki/Logic_Theorist https://en.wikipedia.org/wiki/Logic_Theorist
[2] https://en.wikipedia.org/wiki/Chinook_(computer_program) https://en.wikipedia.org/wiki/Chinook_(computer_program)
[3] https://en.wikipedia.org/wiki/Deep_Blue_versus_Garry_Kasparov https://en.wikipedia.org/wiki/Deep_Blue_versus_Garry_Kasparo...
[4] https://en.wikipedia.org/wiki/Mycin https://en.wikipedia.org/wiki/Mycin
[5] https://github.com/stassa/louise/ https://github.com/stassa/louise/
- Veedrac 6y ago> You can consult wikipedia or e.g. the calls for papers from major AI conferences, AAAI and IJCAI, if in doubt. As to Wikipedia, see the second paragraph. The sections where it mentions, eg., symbolic or sub-symbolic approaches are prefixed with “Researchers in the 1960s and the 1970s” or “By the 1980s”. Kind'a telling. Like, my point is not about whether you can find the odd person trying to solve intelligence with grammars, or what were GOFAI conferences still harbour GOFAI research in the corners, my point is that a) these approaches don't work as a way to actually tackle AI, the problem, b) the vast majority of the field does not take them as seriously as a method of doing so, regardless of other uses, and c) therefore it's natural, not ‘impossible’, to gain AI expertise without having much care for those parts of the field. > Further, I can certainly point you to successes of symbolic AI, which you say don't exist. For one thing, the entire fields of automated theorem proving, planning, search, game playing, knowledge representation and reasoning, etc. that you say are "not AI", but are like I say still active and still state of the art in their respective tasks. Yes, but there's a reason I suffixed that comment with “(again, in an AI context)”. GOFAI is great if you ignore the last two letters of the name, and how it failed almost all its major promises. These used to be considered AI because it was thought that you could build a useful reasoning agent out of a combination of these techniques, given appropriate developments. Now (almost) nobody does that; Google Map's pathfinding is just a pathfinder, not a general reasoner. > Below I've defined the problem in the format expected by Louise Right, OK, I figured it'd be something like this, when you said it was trivial, but this is just another perspective on my original criticism. You wrote the program you wanted it to generate as the background knowledge. It must be so, because your examples don't specify, even roughly approximately, the program you wanted generated. Another valid solution would be (among many) ordered(A,B,C):-shorter(A,B),shorter(A,C). and the only reason it didn't choose this is because you gave it the program you wanted it to generate (obfuscated a little, yet still, there was only one application available). It didn't ‘learn’ anything. > I should point out that there exists no neural net approach that can learn the same (or a similar) program from a single positive example- not least because neural nets cannot make use of background knowledge (i.e. a library of programs from which to build other programs). This is not true, but as you wanted to prune, I'll leave it there.
- 6y ago