AbstractThe human brain is one of the most complex systems known to mankind with an estimated 100 billion neurons, 0.15 quadrillion (150 thousand billion) synapses and around a trillion glial cells. Understanding this intricate organism has long been the objective of many scientists and researchers. The motivations for comprehending the human brain can be out of sheer curiosity, to cure diseases (e.g. Alzheimer’s and depression) or to build efficient braininspired computers. More recently, progress has been achieved in modelling the role of glia cells, a cell that was not previously recognised as a key player in brain repair. To this end, mathematical models have been developed for predicting the behaviour of cells (neuronal and glia) under stimuli. The models range in complexity from highly rich in detail such as Hodgkin-Huxley to simpler models like Leaky-integrate and Fire (LIF). An artificial neural network is a grouping of neurons in a certain structure such as feed forward. Spiking Neural Networks (SNNs) are the third generation of neural networks and employ the temporal computing abilities of the human brain. Programming languages such as Matlab or Pynn are used to simulate SNNs however, simulations exhibit signification execution times for large networks. The inclusion of astrocytes, a type of glia cell, - in an SNN provides Spiking Astrocyte Neural Network (SANN). Due to the high computational and complexity of astrocytes, SANNs suffer from even further increased execution times.
This PhD research addresses the issue of prolonged simulation times by utilizing dedicated hardware on FPGAs for accelerating SANN simulations. Astrocytes are incorporated into the simulated models to give the SANN the ability to self-repair. Since FPGAs have limited resources, a NoC-based multiFPGA infrastructure is developed to accommodate SANNs with resource demands exceeding a single FPGA. On top of that, a monitoring and configuration platform is implemented to configure various aspects of the network at the start of operation and to take data off-chip for storage and analysis on a PC during simulations.
The following points summarize the major contributions of this study;
1- Developing a new 32-bit fixed-point hardware prototype for biologically faithful astrocyte model.
2- Incorporating the astrocyte prototype with hardware models of neurons and synapses to facilitate a self-repairing FPGA-based SANN Accelerator (FSA).
3- Designing a novel NoC router, NoC infrastructure and data format to facilitate scalable a multi-FPGA NoC AstroByte platform.
4- A novel FPGA Configuration and Monitoring Platform (FCMP) was utilized for injecting faults, configuring the AstroByte platform and capturing real time simulation data for monitoring and analysis on a PC.
|Date of Award||Jul 2020|
|Sponsors||Vice Chancellor's Research Scholarship (VCRS)|
|Supervisor||Bryan Gardiner (Supervisor), Liam Mc Daid (Supervisor) & Jim Harkin (Supervisor)|
- System-on-Chip (SoC)
- Data Acquisition
- Network-on-Chip (NoC)
- Fault tolerance
- FPGA Acceleration
AstroByte: Programmable and Scalable Multi-FPGA Interconnect infrastructure for Accelerated Simulations of Self-Repairing Spiking Astrocyte Neural Networks
Haji Karim, S. (Author). Jul 2020
Student thesis: Doctoral Thesis