A general complex neural unit integrated with synaptic connection and lateral interaction is proposed. Various kinds of connections can be generated automatically by using different routing codes contained in the input stimuli. Based on the mixed negabinary system, an inner-product algorithm is suggested to perform complex matrix-vector multiplications without carries, signs and decimal points. An incoherent optical correlator with spatial coding is developed to execute the algorithm, thus the complex neural network is implemented optically in parallel. Numerical simulations and experiments of laterally inhibited Hopfield neural network are demonstrated.