Re: Can we neural networks to predict code change?

joshipura@xxxxxxxxx wrote:
If it can predict imminent *place* (forget content) of change >50% of
time, it could be tremendously helpful.

That means huge gains in being able to predict our release cycles and
marketing and even customer support.

All we need is an big enough, properly trained ANN that predicts how a
graph evolves from time to time.

Has anyone done this?


My question:
> Is it theoretically possible to
> A. Train a neural network about past and current code in form of change
> in graph and
> B. Ask it to predict what the next change could be?

-Bhu***



A & B. The question is accuracy or how close is the approximation.
The ANN training rule(s) have a bias, even the so called self-learning
which means the range(s) out of the possible domain can vary.

The ANNs as mentioned before are TM equivalent, because the human
brain neural networks are not the same as ANNs. As you know the brain
receives huge amounts of sensory data. It filters out most of this and
presents to our consciousness what evolution (or maybe God) has deemed
"essential" to our survival. This information is integrated by our
brains/minds to produce survival strategies-- I mean our reasoning
operates on data which has a particular bias (which is going to exclude
some choices) and if this is not uncomputable, it is intractable. The
random factors in the evolution to evolve consciousness are many times
larger than the atoms in the universe, so no computer simulation. And
if you believe in God, the mind of God is inscrutable to humans. Also
I think the instinct to survive is linked to at least a primitive sense
of self-awareness (cockroach) up to humans, which introduces the next
topic:

However, there is quite a bit of progress in "self-aware" networks
(Minsky deprecates this terminology) which might interest you. These
might fall under the category of Random Neural Networks (RNNs) so
to answer your question, a qualified yes, especially with routers
and network traffic which does self-healing load balancing which
requires some prediction. There is a theorem however, which states
that no program can verify all other complex programs (loosely).
Here is another fundamental limitation from John Case and COLT.

http://www.cis.udel.edu/~case/colt.html

"Consider the problem of finding a rule for generating a sequence of numbers such as 9, 61, 52, 63, 94, 46, 18, 1, 121, 441, ... . Here is a rule for this sequence. First compute the squares of successive integers beginning with 3, but, then, to generate the sequence, use, in place of these squares, the squares each with its decimal digits written down in reverse order (ignoring any lead zeros). N.B. This rule can be written as a formal algorithm (or computer program). The problem of finding such rules gets harder as the sequences to generate get more complicated than the one above. Can the rule finding itself be done by some computer program? Interestingly, it is mathematically proven that there can be no computer program which can eventually find (synonym: learn) these (algorithmic) rules for all sequences which have such rules!

Here is a different problem. Given pictures of animals each (correctly) labeled as being or not being a picture of a bird, eventually find/learn an (algorithmic) rule for deciding, of each of those (past, present, and future) pictures, whether or not it depicts a bird. People seemingly solve this problem in childhood as well as many other problems of learning (subconscious) rules to predict membership in concepts (such as the concept of bird). Many cognitive scientists seek to model all of cognition (including concept learning) by computer program. It is, then, an interesting, very hard problem to write a computer program which can learn (algorithmic) rules to predict membership in all the concepts that people can . It's so hard we don't know how to do it yet, and we don't know how to to prove it can't be done."

SH: This includes (computer) neural networks; also these paragraphs
don't claim that a human can surpass (or equal) the computer's ability.
So "code change" not in a general sense.

--------------------------------------------------------------
Java and Networks, which employ the term self-aware in
describing software and frequently mention biological inspiration
for this this type of architecture; the use self* terms including
self-awareness is well-established and worldwide.

------------------------------------------------------------

Melanie Mitchell:
http://www.cs.pdx.edu/~mm/self-awareness.pdf 2005 (page 3) ...

"What is Really Meant by "Self-Awareness" in These Systems?"

"Some readers may disagree with my use of the term "self-awareness"
to describe the kinds of adaptive feedback mechanisms sketched
above. A devil's advocate might argue that "self-regulation via
multiple feedback systems" would be a better description, since
the immune system and ant colonies are not really conscious in
the way the brain is. Perhaps this is true (and I think it
probably is) but then the devil's advocate must explain, precisely,
what the difference is. In the first paragraph of this paper I
defined self-awareness as information contained in a system about
its global state that feeds back to adaptively control the system's
low-level components. Given this definition, both the immune
system and ant colonies have some degree of self-awareness. It is
not clear where, exactly, the "self" is located in these systems,
but neither is it clear for the brain. If there is something more
to self-awareness in the brain, it needs to be elucidated.
Some of the most useful discussions I have seen along these lines
are in the works of Hofstadter and Dennett (Hofstadter 1979;
Hofstadter & Dennett 1981; Dennett 1991). Obviously, these issues
are worth exploring for anyone interested in metacognition."

Regards,
Stephen
.

Quantcast