Neuromorphic event-based image sensors capture only the dynamic motion in a scene, which is then transferred to computation units for motion recognition. This approach, however, leads to time latency and can be power consuming. Here we report computational event-driven vision sensors that capture and directly convert dynamic motion into programmable, sparse and informative spiking signals. The sensors can be used to form a spiking neural network for motion recognition. Each individual vision sensor consists of two parallel photodiodes with opposite polarities and has a temporal resolution of 5 μs. In response to changes in light intensity, the sensors generate spiking signals with diferent amplitudes and polarities by electrically programming their individual photoresponsivity. The non-volatile and multilevel photoresponsivity of the vision sensors can emulate synaptic weights and can be used to create an in-sensor spiking neural network. Our computational event-driven vision sensor approach eliminates redundant data during the sensing process, as well as the need for data transfer between sensors and computation units.