Abstract
Neural systems use the same underlying computational substrate to carry outanalog filtering and signal processing operations, as well as discrete symbolmanipulation and digital computation. Inspired by the computational principlesof canonical cortical microcircuits, we propose a framework for using recurrentspiking neural networks to seamlessly and robustly switch between analog signalprocessing and categorical and discrete computation. We provide theoreticalanalysis and practical neural network design tools to formally determine theconditions for inducing this switch. We demonstrate the robustness of thisframework experimentally with hardware soft Winner-Take-All and mixed-feedbackrecurrent spiking neural networks, implemented by appropriately configuring theanalog neuron and synapse circuits of a mixed-signal neuromorphic processorchip.