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?

-Bhu***

What you are talking about is called a General Purpose Learner or
sometime an Inductive Inference Machine. They all have limitations.

http://www.cis.udel.edu/~case/papers/journal-chen5plus.ps
Costs of General Purpose Learning (with K. Chen and S. Jain

The chances are less than one in a googleplex that a program will
evolve to match the human heuristic. So such an approach, currently
not as good as human judgment, perhaps in the future better than
human judgment, is not going to be implemented identical to human,
by a program, ANN or otherwise. Also if some program is better, it
will very hard for a human to know it, similar to how difficult it
is for human to fathom some of those huge computer assisted proofs.

The current situation is already useful. Learning systems can
benefit from "contexts"-> auxillary learning "related" tasks
which seem to me function like training for ANNs. The challenge
is to find all the rules which contribute to establishing an
overall pattern and to weight them properly in a timely fashion,
which changes learning a given task from impossible to possible.
http://citeseer.ist.psu.edu/case98robust.html
"In Section 4 we have shown that there are very hard learning
problems which become learnable when a suitably selected context
is supplied to the learner." ...

.... if one considers context mappings im­plementable by program
mappings, [Here,] in some cases, the context mapping can even
be implemented by a com­putable program mapping. However, on the
other hand, we also showed that in general there does not exist
an upper bound on the Turing degree which a program mapping may
need to provide useful selected contexts for all tasks in a
particular class." ...

"Hence, a natural question is if it is possible to invent
a learning notion that "avoids all forms of self-­referential
coding" tricks. A little thought suggests that this may
be a bit too much to ask for because if there is to be some
kind of learnability at all, then there is certainly some
kind of coding. So the more interesting question is per­haps
the formulation of a learning model that avoids in­tuitive
/reasonable forms of self-­referential coding tricks.
Is there some hierarchy of more and more sophisticated
coding tricks? And if so, does such a hierarchy inter­act
in any way with the many learnability hierarchies
known in inductive inference? We feel answers to these
questions will improve our understanding of learnability."

Just where does the oracle topologically fit into the matrix,
Stephen
.

Quantcast