Predictive Coding Light: learning compact visual codes by combining excitatory and inhibitory spike timing-dependent plasticity.

It is widely believed that the brain uses predictive coding schemes to represent sensory inputs in an efficient manner. However, it is still debated how networks of spiking neurons can learn such coding schemes in an unsupervised fashion. Here we present a hierarchical spiking neural network architecture that learns an efficient encoding of visual input from an event-based vision sensor by combining excitatory and inhibitory spike timing-dependent plasticity (STDP). The network develops receptive fields and exhibits surround suppression effects reminiscent of biological findings. We show that inhibitory STDP which aims to suppress predictable (and therefore redundant) spikes in neurons strongly reduces neural activity (and therefore energy costs) with only moderate reductions in coding fidelity.