Signal reconstruction from spiking neuron models

Abstract

We describe a method for signal reconstruction from spiking neuron models such as integrate-and-fire or leaky integrate-and-fire neurons. These neural models encode a single analog signal in the timing of asynchronous digital pulses. We show that using only the output firing times of these neurons, we can recover a bandlimited input signal to within machine precision. A major application of this work is for a replacement of conventional analog-to-digital converters in some applications where simpler analog hardware is traded off more complex reconstruction on the part of the subsequent digital processor. Realistic SPICE simulations of CMOS spiking neurons show that accurate reconstruction with more than 12-bit precision can be achieved. The effects of frequency aliasing, noise, and temporal quantization are considered.