bakker@cs.uq.oz.au (Paultje Bakker) (04/18/91)
I am interested in pointers to any articles, researchers, papers, or books that have investigated the extraction of rules from a successfully trained neural network. Does anyone know if this is indeed possible? Can human-readable rules be deduced from the distributed weights and connections of a neural network? Any help would be greatly appreciated. Paul Bakker -- --Paul Bakker -- email: bakker@cs.uq.oz.au --Depts. of Computer Science/ Psychology ------ --University of Queensland ---- -------- --New Holland --- ------------- ------------
jdm5548@tamsun.tamu.edu (James Darrell McCauley) (04/19/91)
In article <886@uqcspe.cs.uq.oz.au>, bakker@cs.uq.oz.au (Paultje Bakker) writes: |> I am interested in pointers to any articles, researchers, papers, |> or books that have investigated the extraction of rules from a |> successfully trained neural network. |> On a similar note, Martin Wildberger of General Physics presented something at the Simulation Multiconference in New Orleans a couple of weeks ago on using weights to determine significance of inputs and then "fuzzyfy" them and change them to verbage so that a non-techie could understand why a NN came to a particular solution. This is a second-hand account of his talk - I was unable to stay in town. Does anyone have any references to this type of thing? This again is second-hand, but I heard that when folks asked for references (or even copies of his slides) that he referred them to a publication last year, either in the SMC Proceedings or some NN conference. -- James Darrell McCauley, Grad Res Asst, Spatial Analysis Lab Dept of Ag Engr, Texas A&M Univ, College Station, TX 77843-2117, USA (jdm5548@diamond.tamu.edu, jdm5548@tamagen.bitnet)
sfp@mars.ornl.gov (Phil Spelt) (04/19/91)
In article <14940@helios.TAMU.EDU> jdm5548@tamsun.tamu.edu (James Darrell McCauley) writes: >In article <886@uqcspe.cs.uq.oz.au>, bakker@cs.uq.oz.au (Paultje Bakker) writes: >|> I am interested in pointers to any articles, researchers, papers, >|> or books that have investigated the extraction of rules from a >|> successfully trained neural network. >|> > >On a similar note, Martin Wildberger of General Physics presented something >at the Simulation Multiconference in New Orleans a couple of weeks ago on >using weights to determine significance of inputs and then "fuzzyfy" them >and change them to verbage so that a non-techie could understand why a NN >came to a particular solution. > >This is a second-hand account of his talk - I was unable to stay in town. >Does anyone have any references to this type of thing? This again is >second-hand, but I heard that when folks asked for references (or even copies >of his slides) that he referred them to a publication last year, either >in the SMC Proceedings or some NN conference. > >-- >James Darrell McCauley, Grad Res Asst, Spatial Analysis Lab >Dept of Ag Engr, Texas A&M Univ, College Station, TX 77843-2117, USA >(jdm5548@diamond.tamu.edu, jdm5548@tamagen.bitnet) I have seen several postings on this topic in this newsgroup, so I fianlly decided that no-one reading this group is aware of work being done at FL State Univ. The work started as graduate work by Dave Kuncicky ("kun- sisky"), and has bee picked up by others in the FSU CS Department: Susan Hruska and Chris Lacher, most noteably. Their "mission" is to explore the transfer of knowledge between [simple] exeprt systems and neural nets -- in both directions. Snail-mail for these people is: Department of Computer Science Florida Statre University Tallahassee, FL 32306 Their work has been presented at both the Auburn Workshops on ANNs ('90 and '91). Although I was skeptical a year ago about the reults, I have since become convinced that this is a potentially very useful line of investigation -- permitting the "fine tuning" of expert systems by training a specially-designed [backprop] net, then transferring the knowledge back to the ES. The net would be designed on the basis of expert nowledge initialy encoded into the ES. Contact these reserachers if you are interested. They LOVE to talk about their work! ============================================================================= MIND. A mysterious form of matter secreted by the brain. Its chief activity consists in the endeavor to asscertain its own nature, the futility of the attempt being due to the fact that it has nothing but itself to know itself with. -- Ambrose Bierce ============================================================================= Phil Spelt, Cognitive Systems & Human Factors Group sfp@epm.ornl.gov ============================================================================ Any opinions expressed or implied are my own, IF I choose to own up to them. ============================================================================
guedalia@bimacs.BITNET (David Guedalia) (04/21/91)
In article <886@uqcspe.cs.uq.oz.au> bakker@cs.uq.oz.au writes: >I am interested in pointers to any articles, researchers, papers, >or books that have investigated the extraction of rules from a >successfully trained neural network. > >Does anyone know if this is indeed possible? Can human-readable >rules be deduced from the distributed weights and connections of a >neural network? > What type of network are you talking about. I remember some mention on this board about using a neural net. in a expert system, would that be the same? In a Kohonen feature map the weights would not say much. But the distribution of the weights in the map should have some meaning. Has anyone heard or have any ideas about how one could represent a feature map not by its weights but by the relationship between its neighborhoods ? I have seen something called instars and out-stars an out-star could be a feature map and an instar would be the oppositte a way of representing the feature map by a single vector. Has anyone seen refrences to that? david