Date of Award

August 2017

Degree Type


Degree Name

Doctor of Philosophy (PhD)


Electrical Engineering and Computer Science


Qinru Qiu


Bayesian inference, Digital neuron, High performance computing, Neuromorphic computing, Simulation, Spiking neural networks

Subject Categories



Neuromorphic computing refers to brain inspired algorithms and architectures. This paradigm of computing can solve complex problems which were not possible with traditional computing methods. This is because such implementations learn to identify the required features and classify them based on its training, akin to how brains function. This task involves performing computation on large quantities of data. With this inspiration, a comprehensive multi-pronged approach is employed to study and efficiently implement neuromorphic inference model using heterogeneous clusters to address the problem using traditional Von Neumann architectures and by developing spiking neural networks (SNN) for native and ultra-low power implementation. In this regard, an extendable high-performance computing (HPC) framework and optimizations are proposed for heterogeneous clusters to modularize complex neuromorphic applications in a distributed manner. To achieve best possible throughput and load balancing for such modularized architectures a set of algorithms are proposed to suggest the optimal mapping of different modules as an asynchronous pipeline to the available cluster resources while considering the complex data dependencies between stages. On the other hand, SNNs are more biologically plausible and can achieve ultra-low power implementation due to its sparse spike based communication, which is possible with emerging non-Von Neumann computing platforms. As a significant progress in this direction, spiking neuron models capable of distributed online learning are proposed. A high performance SNN simulator (SpNSim) is developed for simulation of large scale mixed neuron model networks. An accompanying digital hardware neuron RTL is also proposed for efficient real time implementation of SNNs capable of online learning. Finally, a methodology for mapping probabilistic graphical model to off-the-shelf neurosynaptic processor (IBM TrueNorth) as a stochastic SNN is presented with ultra-low power consumption.


Open Access

Included in

Engineering Commons