omlinc@cs.rpi.edu (Christian Omlin) (04/26/91)
Hi ! I am running simulations with backprop networks. The network is used as a classifier. I am interested in the sensitivity of the network to perturbations in the weights. My experiments indicate that the performance degrades more rapidly when the weights from the input to the hidden layer are perturbed as opposed to perturbation of weights from the hidden to the output layer. This implies that, for my experiments, the shape of the decision regions is largely determined by the first hidden layer. Are there any references (simulations, etc) confirming this behavior ? Thanks. Christian ---------------------------------------------------------------------------- Christian W. Omlin office: home: Computer Science Department Foxberry Farm Amos Eaton 119 Box 332, Route #3 Rensselaer Polytechnic Institute Averill Park, NY 12018 Troy, NY 12180 USA (518) 766-5790 (518) 276-2930 e-mail: omlinc@turing.cs.rpi.edu ----------------------------------------------------------------------------
aam9n@helga2.acc.Virginia.EDU (04/29/91)
In article <j+wg+7.@rpi.edu> omlinc@cs.rpi.edu (Christian Omlin) writes: >Hi ! > >I am running simulations with backprop networks. The network is used as >a classifier. >I am interested in the sensitivity of the network to perturbations >in the weights. My experiments indicate that the performance degrades >more rapidly when the weights from the input to the hidden layer are >perturbed as opposed to perturbation of weights from the hidden to >the output layer. This implies that, for my experiments, the shape >of the decision regions is largely determined by the first hidden >layer. Are there any references (simulations, etc) confirming this >behavior ? The most directly relevant paper for this would be: M. Stevenson, R. Winter & B. Widrow, "Sensitivity of Feedforward Neural Networks to Weight Errors", IEEE Trans. on Neural Networks, vol. 1, no. 1, pp. 71-80, March 1990. I am currently working on the robustness of feed-forward nets with real-valued outputs, but I am looking at perturbations in neuron outputs. I have done what I think is a fairly thorough literature search, but would appreciate any references, pointers etc. that might address this issue. If there is interest, I will summarize to the net. Thanks. Any takers for a discussion of neural net fault-tolerance? Regards, Ali Minai University of Virginia aam9n@Virginia.EDU