home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!wupost!sdd.hp.com!hplabs!ucbvax!NEURON.SIEMENS.COM!kpfleger
- From: kpfleger@NEURON.SIEMENS.COM (Karl Pfleger)
- Newsgroups: comp.ai.neural-nets
- Subject: Re: Wild values (was Reducing Training time ...)
- Message-ID: <9208201551.AA02766@neuron.siemens.com>
- Date: 20 Aug 92 15:51:31 GMT
- Sender: daemon@ucbvax.BERKELEY.EDU
- Lines: 50
-
- Newsgroups: comp.ai.neural-nets
- Subject: Re: Wild values (was Reducing Training time ...)
- Summary:
- Expires:
- References: <arms.714146123@spedden> <36967@sdcc12.ucsd.edu> <arms.714208873@spedden>
- Sender:
- Followup-To:
- Distribution:
- Keywords: back propagation, training, generalisation
-
- 1 quick point and 1 wild idea:
-
- First, if one desires to avoid wild output values for certain regions of
- input space, one ought to have training pairs from that region of input
- space in the training set. The point about desiring certain behavior on
- 0 to 1 and not including any training pairs from that region has already
- been made.
-
- Vaguely similar: since the inputs that will be thrown at the system in actual
- use will have some probability distribution based on whatever the system is
- doing, the training set should be generated by sampling the same
- distribution, or something as close to it as possible, NOT by picking a
- few values by hand or by using a lattice or points regularly spaced
- (unless that represents the distribution well).
-
- I have a much more difficult time picturing wild values coming from a
- network trained on a significant number of random, real inputs than I do
- coming from a network trained on a handful of regularly spaced
- integers.
-
-
- A wild idea for people trying to avoid wild values (e.g. for safety
- critical applications etc.): Once the network has been trained and the
- weights are fixed, it should be possible to determine the maximum and
- minimum output values for all inputs. This can be done with normal
- gradient ascent and descent. Simply calculate partials of the output(s)
- with respect to the input(s). Convergence should be much faster than
- training networks in the first place due to the generally smaller number
- of inputs than weights. Multiple runs from different starting positions
- or the use of stochastic techniques likely to converge to global
- maxima/minima can reduce the chance of not seeing a wild value that
- actually exists. With a small number of inputs (definitely 1, maybe
- a more) analytical techniques should be able to provably determine the
- global maximum and minimum.
-
- This idea's usefulness is limited by a number of requirements such as
- fixed weights and wild values meaning large or small, but still seems
- like it should have fairly wide applicability.
-
- -Karl kpfleger@learning.siemens.com
-