V. COMPARISON WITH EXISTING METHODS

In neural networks with complex neurons, the angle between adja-cent complex points on the unit circle is inversely proportional to the number of gray levels. As a consequence, the recall performance de-teriorates as the number of gray levels increases. For example, in [5] and [6] the number of gray levels used in the design examples are 8 and 20, respectively. The same problem occurs using multivalued acti-vation functions with normalized output range (e.g., between 1 and +1 as in [2]). In the proposed method, the recall performance is satis-factory with the commonly used number of gray levels, i.e., 256. A second aspect is the simple implementation. The learning algo-rithm described in Section III is completely local in the sense that the weights of each neuron can be computed independently from the others; so the learning algorithm is suitable to parallel computation. Moreover, it requires only add operations and a few bits of digital pre-cision. The same advantages are shared by the network described i