Continuous Time Spiking Neural Network Field Programmable Neural Network Array

Digital signal processing (DSP) uses digital computers to process sensor signals from physical systems, which are usually described with real-valued differential equations in continuous time.   Modern digital computers, however, represent numerical quantities with integers and process data synchronously in a series of discrete operations.  Traditional DSP sampling rounds the signal value to the nearest integer at a fixed sampling frequency, which introduces quantization noise and limits the frequencies that can be represented without error.  Continuous-time DSP (CT-DSP), however, instead uses level crossing sampling to generate and process samples each time the signal crosses a threshold, leading to a reduction in quantization error, representation of arbitrarily fast signals, and power savings.  One emerging application in which CT-DSP may yield significant benefits is in spiking neural networks (SNNs).  SNNs are neuromorphic computing devices that asynchronously process a series of spikes with artificial neurons, and they typically encode inputs in terms of differences, which is well aligned to level-crossing sampling.