Experimental evidence suggests that the visual recognition of biological movements is based on learned spatio-temporal templates. Work in computational vision shows that movement recognition can be accomplished by recognizing temporal sequences of form or optic flow patterns. Recurrent
neural networks with asymmetric lateral connections are one physiologically
plausible way for the encoding of spatio-temporal templates. We demonstrate
that time-dependent hebbian plasticity is suitable for establishing the required
lateral connectivity patterns. We tested different hebbian plasticity rules and compared their efficiency and stability properties in simulations and by mathematical analysis. We found the most robust behavior for a learning rule that assumes a normalization of the total afferent synaptic connectivity that can be supported by each neuron. Consistent with psychophysical data our model learns the appropriate lateral connections after less than 30 stimulus repetitions. The resulting recurrent neural network shows strong sequence selectivity.