Training Feed Forward Nets with Binary Weights via a Modified CHIR Algorithm
D. Saad
E. Marom
Faculty of Engineering, Tel Aviv University,
Ramat Aviv 69978, Israel
Abstract
Learning by choice of internal representations (CHIR) is a learning algorithm for a multilayer neural network system, introduced by Grossman et al. [2,3], based upon determining the internal representations of the system as well as its internal weights. In a former paper [8] we have shown a method for deriving the CHIR algorithm whereby the internal representations (IR) as well as the weights are allowed to be modified via energy minimization consideration. This method is now applied for training a feedforward net with binary weights, supplying a convenient tool for training such a net. Computer simulations show a fast training process for this algorithm in comparison with backpropagation [7] and the CHIR [2,3] algorithms, both used in conjunction with a feedforward net with continuous weights. These simulations include the restricted cases of parity, symmetry, and parity--symmetry problems.
