PhD defence G. (Georgios) Smaragdos

Below is a brief summary of the dissertation:

Among the various methods in neuroscience for understanding brain function, in-silico simulations have been gaining popularity. 
Advances in neuroscience and engineering led to the creation of mathematical models of networks that do not simply mimic biological behaviour in an abstract fashion, but emulate it in significant detail.
Such an example is the Spiking Neural Network (SNN) that can model a variety of additional behavioural features, and can be potentially used to study a wide variety of brain functions.  
In-silico simulations of SNNs can have beneficial results not only for neuroscience research but breakthroughs can also potentially benefit medical, computing and A.I. research. However, SNNs come with computationally demanding workloads that traditional computers might struggle to handle. Thus, the use of High Performance Computing (HPC) is desirable with such applications.

This dissertation explores the topic of HPC-based in-silico brain simulations. Initially, it explores the potential use of FPGAs, which are custom digital chips, for executing SNNs. It later extends this exploration to GPU and multi-core devices. It concludes with the proposal of the BrainFrame system. The proof-of-concept design supports standard and extended Hodgkin-Huxley SNN models. The system integrates a GPU, CPU and FPGA devices while also using a familiar neuroscientific language, PyNN. It can score best-in-class performance while being more user-friendly for the typical model developer. BrainFrame provides both a powerful platform for acceleration and also a front-end familiar to the neuroscientist.