Preview

PDF, English Download (11MB) | Terms of use

Abstract

Accelerated neuromorphic hardware allows the emulation of spiking neural networks with a high speed-up factor compared to classical computer simulation approaches. However, realizing a high degree of versatility and configurability in the implemented models is challenging. In this thesis, we present two mixed-signal ASICs that improve upon previous architectures by augmenting the versatility of the modeled synapses and neurons. In the first part, we present the integration of an analog multi-compartment neuron model into the Multi-Compartment Chip. We characterize the properties of this neuron model and describe methods to compensate for deviations from ideal behavior introduced by the physical implementation. The implemented features of the multi-compartment neurons are demonstrated with a compact prototype setup. In the second part, the integration of a general-purpose microprocessor with analog models of neurons and synapses is described. This allows to define learning rules that go beyond spike-timing dependent plasticity in software without decreasing the speed-up of the underlying network emulation. In the third part, the importance of testability and pre-tapeout verification is discussed and exemplified by the design process of both chips.