Vorschau

PDF, Englisch Download (14MB) | Nutzungsbedingungen

Abstract

Mixed-signal accelerated neuromorphic hardware is a class of devices that implements physical models of neural networks in dedicated analog and digital circuits. These devices offer the advantages of high acceleration and energy efficiency for the emulation of spiking neural networks but pose constraints in form of device variability and of limited connectivity and bandwidth. We address these constraints using two complementary approaches: At the network level, the influence of multiple distortion mechanisms on two benchmark models is analyzed and compensation methods are developed that counteract the resulting effects. The compensation methods are validated using a simulation of the BrainScaleS neuromorphic hardware system. At the single neuron level, calibration procedures are presented that counteract device variability for a new analog implementation of an adaptive exponential integrate-and-fire neuron model in a 65 nm process. The functionality of the neuron circuit together with these calibration methods is verified in detailed transistor-level simulations before production. The versatility of the circuit design that includes novel multi-compartment and plateau-potential features is demonstrated in use cases inspired by biology and machine learning.