Imaging black boxes with multiple stochastic output
IEEE Transactions on Systems, Man and Cybernetics
Internal structure of black boxes with multiple stochastic output
Signal Processing - Multidimensional Signal Processing, Part II
Learning and relearning in Boltzmann machines
Parallel distributed processing: explorations in the microstructure of cognition, vol. 1
| Hi-index | 0.00 |
The performance of neural networks often is characterized by the percentage of correct responses. In this study it is advocated to use the more sensitive and rigorous methods of signal detection theory. According to this theory the sensitivity of the network and the detectability of patterns in noise are obtained in much the same way as has been practiced in psychophysics for years. Moreover, by using concepts of detection theory a method is derived by which the internal structure of a neural net can be depicted in a so called black box image. Although the methods proposed here apply quite generally to different kinds of network, the theory was demonstrated by application to a particular kind of back propagation network. This network had a 5 x 5 input array with patterns of letters distorted by noise. The number of different letters and the number of hidden units were used as experimental parameters. The sensitivity of the network was measured at all stages during the learning process. The method appears to provide indices of performance for various levels of details.