Many connectionist and neural network models of cognition work on the assumption that information is transmitted between the processing nodes (neurons) using continuous activation levels. Translated into neurophysiological terms, this effectively considers that neurons use a rate-based coding scheme, often using Poisson-like activation. I will review evidence that this form of coding is incompatible with the speed with which complex natural scenes can be processed. An alternative coding scheme uses the fact that the most strongly activated neurons will tend to fire first, and, as a consequence, information can be extracted from the order in which neurons within a population fire. I will illustrate the power of such an approach by showing that spike-time dependent plasticity, when coupled with temporally structured firing, automatically leads to high weights being concentrated on the earliest firing inputs. We have recently shown that such a mechanism will naturally result in neurons that respond selectively and rapidly to frequently occurring input patterns.