Event-based computation: unsupervised elementary motion decomposition

Fast, localised motion detection is crucial for an efficient attention mechanism. We show that modelling a network capable of such motion detection can be performed using spiking neural networks simulated on many-core neuromorphic hard-ware. Moreover, highly sensitive neurons arise from the presented network architecture through unsupervised self-organisation. We use a synaptic rewiring rule which has been shown to enable the formation and refinement of neural topographic maps. Our extension allows newly formed synapses to be initialised with a delay drawn from a uniform distribution. Repeated exposure to moving bars enables neurons to be sensitised to a preferred direction of movement. Incorporating heterogeneous delays results in more sensitive neural responses. A readout mechanism involving a neuron for each learnt motion is sufficient to establish the input stimulus class.