home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!munnari.oz.au!bunyip.cc.uq.oz.au!uqcspe!cs.uq.oz.au!mav
- From: mav@cs.uq.oz.au (Simon Dennis)
- Newsgroups: comp.ai.neural-nets
- Subject: Re: NNs that learn to learn
- Message-ID: <11417@uqcspe.cs.uq.oz.au>
- Date: 14 Dec 92 00:48:21 GMT
- References: <7240582678-313510@exaib>
- Sender: news@cs.uq.oz.au
- Reply-To: mav@cs.uq.oz.au
- Lines: 37
-
- bruhn@exaib (Peter Bruhn ) writes:
-
- >So my question is: are there any NNs that adapt to changes in their
- >environments? So the underlying procedure should look like:
-
- > 1. They are trained
- > 2. They are used and adapt automatically to
- > changes in their environments
-
- >Peter (bruhn at uxe.wu-wien.ac.at)
-
- Learning to learn phenomena are well documented in the human memory literature.
- Subject's ability to "memorize" or "learn" to recall lists of items
- improves as they are exposed to multiple trials. In contrast, subject's ability
- to do episodic recognition (i.e. Was this word in the list you just saw?) does
- not seem to get any better (at least after about 5 years of age). It is
- difficult in the second case however to determine whether subjects have
- just reached ceiling.
-
- There have been a couple of attempts to include models from the human memory
- literature into backprop networks. At NIPS, Tony Plate presented a model which
- included a derivative of Ben Murdock's TODAM model, and I have done something
- similar with Ray Pike's matrix memory. During testing, the backprop weights
- are held constant, but the "weights" of the matrix memory are allowed to
- change, so as to store items and hence adapt to the current environment.
- This not only leads to massive improvements in capacity over normal
- recurrent network architectures, but also allows the modelling of
- learning to learn, or perhaps more accurately, learning to memorize
- phenomena.
-
- Simon.
-
- --
- Simon Dennis Address: Department of Computer Science
- Email: mav@cs.uq.oz.au University of Queensland
- QLD 4072
- Australia
-