Digital pulse processing

[摘要] This thesis develops an exact approach for processing pulse signals from an integrate-and-fire system directly in the time-domain. Processing is deterministic and built from simple asynchronous finite-state machines that can perform general piecewise-linear operations. The pulses can then be converted back into an analog or fixed-point digital representation through a filter-based reconstruction. Integrate-and-fire is shown to be equivalent to the first-order sigma-delta modulation used in oversampled noise-shaping converters. The encoder circuits are well known and have simple construction using both current and next-generation technologies. Processing in the pulse-domain provides many benefits including: lower area and power consumption, error tolerance, signal serialization and simple conversion for mixed-signal applications. To study these systems, discrete-event simulation software and an FPGA hardware platform are developed. Many applications of pulse-processing are explored including filtering and signal processing, solving differential equations, optimization, the minsum / Viterbi algorithm, and the decoding of low-density parity-check codes (LDPC). These applications often match the performance of ideal continuous-time analog systems but only require simple digital hardware. Keywords: time-encoding, spike processing, neuromorphic engineering, bit-stream, delta-sigma, sigma-delta converters, binary-valued continuous-time, relaxation-oscillators.

[发布机构] Massachusetts Institute of Technology