Noise sensitivity of projection neural networks

The authors analyze network sensitivity to noise introduced at both the neural and interconnect levels. Two ANNs (artificial neural networks) are analyzed. The first is the alternating projection neural network (APNN), which can be used as a content-addressable memory. The authors show that its performance improves as the percentage of clamped neurons is increased. The second is the layered classifier artificial neural network (L-CANN), whose performance is based on nonlinearly augmented synthetic discriminant functions (also called composite matched filters). It is shown for the L-CANN that the covariances are strongly dependent on the number of hidden neurons. For both ANNs, if the signal-to-noise ratio of the interconnects remains constant, the performance improves as the number of hidden neurons is increased