Encoding strategies in multilayer neural networks
E. Elizalde, Sergio Gómez, August Romeo · Journal of Physics A Mathematical and General · 1991
Neural networks capable of encoding sets of patterns are analysed. Solutions are found by theoretical treatment instead of by supervised learning. The behaviour of 2 R (R in N) input units is studied and its characteristic features are discussed. The accessibilities for non-spurious patterns are calculated by analytic methods. Although thermal noise may induce wrong encoding, the authors show how it can rid the output of spurious sequences. Further, they compute error bounds at finite temperature.