Preview

PDF, English Download (13MB) | Terms of use

Abstract

This thesis establishes a scalable multi-user workflow for the operation of a highly configurable, large-scale neuromorphic hardware platform. The resulting software framework provides unified low-level as well as parallel high-level access. The latter is realized by an efficient abstract neural network description library, an automated translation of networks into hardware specific configurations and an experiment server infrastructure responsible for scheduling and executing experiments. Scalability, manual guidance and a broad support for handling hardware imper- fections render the model translation process suitable for large networks as well as large-scale neuromorphic systems. Networks with local connectivity, random networks and cortical column models are explored to study the topological aptitude of the neuromorphic platform and to benchmark the workflow. Depending on the model, performance improvements of more than two orders of magnitude have been achieved over a previous implementation. Additionally, an automated defect assessment for hardware synapses is introduced, indicating that most synapses are available for model emulation. In a second study, a tempotron-based hardware liquid state machine has been developed and applied to different tasks, including a memory challenge and digit recognition. The trained tempotron inherently compensates for fixed pattern variations making the setup suitable for analog neuromorphic hardware. The achieved performance is comparable to reference software simulations.