Form Versus Function: Theory and Models for Neuronal Substrates

The quest for endowing form with function represents the fundamental motivation behind all neural network modeling. In this thesis, we discuss various functional neuronal architectures and their implementation in silico, both on conventional computer systems and on neuromorpic devices. Necessarily, such casting to a particular substrate will constrain their form, either by requiring a simplified description of neuronal dynamics and interactions or by imposing physical limitations on important characteristics such as network connectivity or parameter precision. While our main focus lies on the computational properties of the studied models, we augment our discussion with rigorous mathematical formalism. We start by investigating the behavior of point neurons under synaptic bombardment and provide analytical predictions of single-unit and ensemble statistics. These considerations later become useful when moving to the functional network level, where we study the effects of an imperfect physical substrate on the computational properties of several cortical networks. Finally, we return to the single neuron level to discuss a novel interpretation of spiking activity in the context of probabilistic inference through sampling. We provide analytical derivations for the translation of this ``neural sampling'' framework to networks of biologically plausible and hardware-compatible neurons and later take this concept beyond the realm of brain science when we discuss applications in machine learning and analogies to solid-state systems.