home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!zaphod.mps.ohio-state.edu!uakari.primate.wisc.edu!sdd.hp.com!mips!darwin.sura.net!haven.umd.edu!mimsy!afterlife!hcbarth
- From: hcbarth@afterlife.ncsc.mil (Bart Bartholomew)
- Newsgroups: comp.ai.neural-nets
- Subject: Re: need for unique test sets
- Message-ID: <1992Jul22.033633.18040@afterlife.ncsc.mil>
- Date: 22 Jul 92 03:36:33 GMT
- References: <1992Jul19.070433.5896@afterlife.ncsc.mil> <25633@life.ai.mit.edu> <arms.711688181@spedden>
- Organization: The Great Beyond
- Lines: 27
-
-
- Prof Armstrong and I are converging on agreement.
- In particular, his last two points
- a. without a priori knowledge, no one can know what the "correct"
- function is, based on a fixed finite sample.
- and
- b. in general, no set of tests unless they cover the whole space
- can assure that the neural net output will do what is correct
- even if you know what "correct" means. There has to be a proof
- technique somehow.
- These remarks are clearly correct. In practice, if you don't/can't
- know for sure that you have the correct answer, then the best you
- can do is try to tilt the odds in your favor. The more data you have
- available the better, if you have the compute power to work with
- it all. If there are (as I suspect) an infinite number of functions
- that can generate any data set, then you have no guarantee of
- selecting the 'right' one no matter what you do. All you can
- do is hope that you have at least found an equivalent function.
- Then you pays your money and you takes your chances.
- There are no guarantees in life. Even real-life pilots make
- mistakes and drop people out of the air.
- Bart
- --
- "It's not the thing you fling, the fling's the thing." - Chris Stevens
- If there's one thing I just can't stand, it's intolerance.
- *No One* is responsible for my views, I'm a committee. Please do not
- infer that which I do not imply. hcbarth@afterlife.ncsc.mil
-