This paper pioneers the use of neural networks to provide a fast and automatic way to classify lightcurves in massive photometric datasets. As an example, we provide a working neural network that can distinguish microlensing lightcurves from other forms of variability, such as eruptive, pulsating, cataclysmic and eclipsing variable stars. The network has five input neurons, a hidden layer of five neurons and one output neuron. The five input variables for the network are extracted by spectral analysis from the lightcurve datapoints and are optimised for the identification of a single, symmetric, microlensing bump. The output of the network is the posterior probability of microlensing. The committee of neural networks successfully passes tests on noisy data taken by the MACHO collaboration. When used to process 5000 lightcurves on a typical tile towards the bulge, the network cleanly identifies the single microlensing event. When fed with a sub-sample of 36 lightcurves identified by the MACHO collaboration as microlensing, the network corroborates this verdict in the case of 27 events, but classifies the remaining 9 events as other forms of variability. For some of these discrepant events, it looks as though there are secondary bumps or the bump is noisy or not properly contained. Neural networks naturally allow for the possibility of novelty detection -- that is, new or unexpected phenomena which we may want to follow up. The advantages of neural networks for microlensing rate calculations, as well as the future developments of massive variability surveys, are both briefly discussed.