Learning algorithms for neural networks with ternary weights.
Journal
Neural Networks
Journal Volume
1
Journal Issue
Supplement-1
Pages
166-167
Date Issued
1988
Author(s)
Abstract
Learning can be looked at as the optimization of an objective function over the domain which consists of all possible sets of weights. In the problem that we are investigating, the domain is {-1,0,1}S, where S is the number of weights in the network, and the objective function we choose is the same as that of the back propagation algorithm, i.e. sum of squares of errors. We have tried several possible learning algorithms, using the parity problem as the example. They are: (1) cyclic minimization method, (2) a greedy minimization method, (3) simulated annealing, (4) a genetic algorithm, (5) combination of (1) and (4), (6) combination of (2) and (4). In the paper we discuss several methods for learning ternary weight networks, and propose several new methods based on the genetic algorithm and simulated annealing.
Type
journal article