adverb@bucsb.UUCP (Josh Krieger) (04/28/89)
I think it's important to say one last thing about ART: ART is primarily usefull in a statistically non-stationary environment because its learned categories will not erode with the changing input. If your input environment is stationary, then there may be little reason to use the complex machinery behind ART; your vanilla backprop net will work just fine. -- Josh Krieger
myke@gatech.edu (Myke Reynolds) (04/29/89)
In article <2503@bucsb.UUCP> adverb@bucsb.bu.edu (Josh Krieger) writes: >I think it's important to say one last thing about ART: > >ART is primarily usefull in a statistically non-stationary environment >because its learned categories will not erode with the changing input. >If your input environment is stationary, then there may be little reason >to use the complex machinery behind ART; your vanilla backprop net will >work just fine. > BAM is a the stationary version of ART, and blows backprop out of the water in both power and simplicity. Its less than a linear equation solver, but thats enough to out-preform backprop. That backprop is not much worse, is not only wrong, it makes for a skimpy last ditch effort to argue for a model that has no other defense. -- Myke Reynolds School of Information & Computer Science, Georgia Tech, Atlanta GA 30332 uucp: ...!{decvax,hplabs,ncar,purdue,rutgers}!gatech!myke Internet: myke@gatech.edu
kavuri@cb.ecn.purdue.edu (Surya N Kavuri ) (05/01/89)
In article <18583@gatech.edu>, myke@gatech.edu (Myke Reynolds) writes: > BAM is a the stationary version of ART, and blows backprop out of the > water in both power and simplicity. Its less than a linear equation solver, > but thats enough to out-preform backprop. > Myke Reynolds I do not understand what you mean by "power" but if you look at the memory capacity, BAM's look pathetic. I do not speak for BP, but I heard some explanations that the hidden layers serve as feature detectors (4-2-4 decoder) which shows a likeness(intuitive) to pattern classification methods. Surya Kavuri (FIAT LUX) P.S: What I dispise in relation to BP is the apparent tendencies that people have in romanticizing it. (I should say that the problem is not with BP but with its researchers). I have seen sinful explanations to what the hidden units stand for. I have seen claims that they stand for concepts that could be given physical meanings (sic!). These are baseless dreams that people come with. This is a disgrace to the serious scientific community as it indicates a degeneration. BP is not even Steepest gradient approach, strictly speaking. It does minimization of an error measure. (1) There are no measures of its convergence time.
myke@gatech.edu (Myke Reynolds) (05/01/89)
Surya N Kavuri writes: > I do not understand what you mean by "power" but if you look at the > memory capacity, BAM's look pathetic. Its memory capacity is no less than that of a linear filter, and its size is not limited, unlike BP. Since size = memory capacity, its memory capacity is limited only by your implementation of a linear equation solver. If you don't make the obvious step of using a sparse solver, then it will be pathetic. -- Myke Reynolds School of Information & Computer Science, Georgia Tech, Atlanta GA 30332 uucp: ...!{decvax,hplabs,ncar,purdue,rutgers}!gatech!myke Internet: myke@gatech.edu