Optimizing Spiking Neuromorphic Networks for Information Processing

The human brain efficiently processes information by analog integration of inputs and digital, binary communication. This fundamental design is captured in spiking neural network models that aim to harness the brain's processing power and energy efficiency. Within this thesis, we contribute to the manifold optimization of these models for information processing. To that end, we first consider strategies for the quantification of the ability to process information. Second, we optimize our network implementations for efficiency by exploiting the analog emulation of neuro-synaptic dynamics on neuromorphic hardware, which aims to tie on the energy efficiency of its biological archetype by mimicking key architectural principles. The actual optimization for information processing is targeted in a third step by exploiting two orthogonal approaches: Specifically, we utilize gradient-based methods in supervised learning scenarios and, moreover, we deliberately exploit collective dynamics for information processing that emerged under local unsupervised plasticity. In the last step, we consider overarched optimization strategies to cope with constraints imposed by the neuromorphic implementation. In particular, we alleviate the ubiquitous issue of limited synaptic resources via local structural reconfigurations. By considering all of the termed stages, we highlight the potential processing capabilities of spiking neural networks implemented on analog neuromorphic hardware