From sietsma at latcs1.lat.oz.au Thu Jul 30 22:47:20 1992 From: sietsma at latcs1.lat.oz.au (Jocelyn Sietsma Penington) Date: Fri, 31 Jul 92 12:47:20 +1000 Subject: identification vs. authentication problems Message-ID: <9207310247.AA22458@latcs1.lat.oz.au> In my opinion (based on somewhat limited experience) feed-forward ANNs usually develop a 'default output' or default class, which means most new inputs which are different to the training classes will be classified in the same way. So, for example, if a net has been trained to classify A,B,C and D, the hidden layer(s) will develop units sensitive to the characteristics of, say, A, C and D. This allows the training set to be correctly classified, but most completely new inputs will give the output pattern for B. My (kludgy) response is to add another class of training data. This contains more patterns than any of the 'real' classes, and is as diverse and noisy as I can make it. I use the unary output coding convention (target output for the first class is (1,0,0), for the second class is (0,1,0), etc.) and set the target for the 'unknown' class as all zeroes. What are people's responses to this? Does it address Francoise Fogelman's problem of wanting to authenticate, rather than simply classify? From: Bill Treurniet >The same issue arises when classifying speech data with a network. How does >one reject non-speech data when the network has been trained with only clean >speech data? A kludgy method that worked for me to some extent was to [method deleted] >The method assumes that the actual distribution of hidden unit activations >for each class in the training set is unimodal; i.e., there is only one set >of hidden layer "features" that give rise to a particular classification. >There is also the issue of how to set the rejection criterion. These issues >were sufficiently troubling that I do not use this method as a matter of >course. I agree with you that these are troubling issues. In particular, I would expect that the first hidden layer activations could only be unimodal if the domain is linearly separable, as this means a single layer of units has separated it. Jocelyn Sietsma Email: sietsma at LATCS1.LAT.oz.au Address: Materials Research Laboratory Phone: (03) 246 8660 or (03) 479 1057 PO Box 50, Melbourne 3032  From gary at cs.UCSD.EDU Thu Jul 30 18:46:24 1992 From: gary at cs.UCSD.EDU (Gary Cottrell) Date: Thu, 30 Jul 92 15:46:24 PDT Subject: identification vs. authentication problems Message-ID: <9207302246.AA23694@odin.ucsd.edu> you could use a boltzmann machine to model the hidden unit activations. This would represent the distribution of hidden unit activations, rather than the expectations. Gary Cottrell 619-534-6640 Reception: 619-534-6005 FAX: 619-534-7029 Computer Science and Engineering 0114 University of California San Diego La Jolla, Ca. 92093 gary at cs.ucsd.edu (INTERNET) gcottrell at ucsd.edu (BITNET, almost anything) ..!uunet!ucsd!gcottrell (UUCP)  From sietsma at latcs1.lat.oz.au Thu Jul 30 22:47:20 1992 From: sietsma at latcs1.lat.oz.au (Jocelyn Sietsma Penington) Date: Fri, 31 Jul 92 12:47:20 +1000 Subject: identification vs. authentication problems Message-ID: <9207310247.AA22458@latcs1.lat.oz.au> In my opinion (based on somewhat limited experience) feed-forward ANNs usually develop a 'default output' or default class, which means most new inputs which are different to the training classes will be classified in the same way. So, for example, if a net has been trained to classify A,B,C and D, the hidden layer(s) will develop units sensitive to the characteristics of, say, A, C and D. This allows the training set to be correctly classified, but most completely new inputs will give the output pattern for B. My (kludgy) response is to add another class of training data. This contains more patterns than any of the 'real' classes, and is as diverse and noisy as I can make it. I use the unary output coding convention (target output for the first class is (1,0,0), for the second class is (0,1,0), etc.) and set the target for the 'unknown' class as all zeroes. What are people's responses to this? Does it address Francoise Fogelman's problem of wanting to authenticate, rather than simply classify? From: Bill Treurniet >The same issue arises when classifying speech data with a network. How does >one reject non-speech data when the network has been trained with only clean >speech data? A kludgy method that worked for me to some extent was to [method deleted] >The method assumes that the actual distribution of hidden unit activations >for each class in the training set is unimodal; i.e., there is only one set >of hidden layer "features" that give rise to a particular classification. >There is also the issue of how to set the rejection criterion. These issues >were sufficiently troubling that I do not use this method as a matter of >course. I agree with you that these are troubling issues. In particular, I would expect that the first hidden layer activations could only be unimodal if the domain is linearly separable, as this means a single layer of units has separated it. Jocelyn Sietsma Email: sietsma at LATCS1.LAT.oz.au Address: Materials Research Laboratory Phone: (03) 246 8660 or (03) 479 1057 PO Box 50, Melbourne 3032  From gary at cs.UCSD.EDU Thu Jul 30 18:46:24 1992 From: gary at cs.UCSD.EDU (Gary Cottrell) Date: Thu, 30 Jul 92 15:46:24 PDT Subject: identification vs. authentication problems Message-ID: <9207302246.AA23694@odin.ucsd.edu> you could use a boltzmann machine to model the hidden unit activations. This would represent the distribution of hidden unit activations, rather than the expectations. Gary Cottrell 619-534-6640 Reception: 619-534-6005 FAX: 619-534-7029 Computer Science and Engineering 0114 University of California San Diego La Jolla, Ca. 92093 gary at cs.ucsd.edu (INTERNET) gcottrell at ucsd.edu (BITNET, almost anything) ..!uunet!ucsd!gcottrell (UUCP)