Properties of neural computation were studied in two types of neuronal networks: isolated leech ganglia and neuronal cultures of dissociated cortical neurons from neonatal rats. With appropriate experimental set-ups it was possible to obtain a precise description of the spread of excitation induced by specific inputs. The evoked spatio-temporal electrical activity was characterized by large variability and the electrical activity of neurons activated by the same stimulation was found to be statistically independent to a high degree. The variability presumably originates from basic properties of synaptic transmission, which is stochastic in nature. As a consequence, the large variability of the evoked spatio-temporal electrical activity appears to be a general property of neural computation and a typical feature of neuronal assemblies. It is shown, however, that the observed statistical independence of co-activated neurons may be used to reduce the effects of variability by appropriately averaging or pooling the electrical activity.