TY  - GEN
UR  - https://archiv.ub.uni-heidelberg.de/volltextserver/29106/
Y1  - 2021///
CY  - Heidelberg
A1  - Baumbach, Andreas
AV  - public
ID  - heidok29106
TI  - From microscopic dynamics to ensemble behavior in spiking neural networks
N2  - As the end of Moore?s law nears and the energy demand for computing increases the search for alternative means of computation becomes more and more relevant. One natural paragon is the animal brain as one of the only known naturally occurring general-purpose computing systems. While its computing model is largely unexplained, it has been shown that biologically inspired artificial neurons ? subject to high-frequency noise ? can approximately implement Boltzmann machines. In the first part of this thesis we explore such a spike-based Bayesian computing framework, compare its observed dynamics to simpler stochastic samplers and develop an improved Markovian model for single LIF neurons. The second part of the thesis then focuses on ensemble phenomena, where we show that a nearest-neighbor connected lattice shows a qualitatively different phase diagrams for different microscopic neuron models. Nevertheless, we can recover the correct critical exponent ? in all cases. Finally, we show two functional demonstrations, including a representation of quantum states, on the accelerated mixed-signal neuromorphic system BrainScaleS. This paves the way for a better understanding of the capabilities of this computational model.
ER  -