The ternary Adaline

Neural networks can be quite sensitive to weight errors, thus requiring very precise implementation hardware in order to avoid errors in the input-output map of the network due to drift in the stored values of the weights. An alternative to precise analog hardware is to limit the number of values the weights are allowed to assume. This restriction on the weights affects the fundamental capabilities of the Adaline, yet allows for less precise hardware to be used. The ternary Adaline represents a limiting case in that it restricts the values of the weights to one of three values: +1, -1, and 0; it allows for an excitatory connection, an inhibitory connection, or no connection. The effects of this restriction on the number of logic functions implementable by a single Adaline and on the capacity of an Adaline are examined. Two algorithms which can be used to training a single ternary Adaline are also introduced