Interactional parallel processing in neural nets
Porada, Boulanger · 1989
The following problem is solved. For a given sequence of associations (I/sub 1/, Y/sub 1/), (I/sub 2/, Y/sub 2/), . . ., (I/sub K/, Y/sub K/) (which are pairs of polar vectors) find a matrix W which solves the equations I/sub k/*W= rho /sub k/Y/sub k/, k=1, 2, . . ., K, with positive coefficients rho k. The solution is found in an explicit form and the condition for the existence of W is stated. If the matrix W is implemented as the weight matrix in a two-layer neural network, the associations (I/sub k/, Y/sub k/) are executed infallibly; the respective computer simulation is described. The algorithm producing W can be implemented in the form of a self-programming of the network, but a layer of hidden neurons is postulated. The three-layer model and the process of self-programming are presented.>