Neuromorphic Computing in Extreme Environments

Supervisors:

Primary supervisor A/Prof. Gregory Cohen

Description:

Neuroscience started with the advent of firearms. The study of gun head wounds led to the realization that the brain is organized in regions responsible for specific functions. It also showed that a brain can function even in the presence of local lesions.

Modern computers do not share this resilience. A single bit flip or a single contact failure can break the entire machine. Fortunately, decades of research and development in Very Large-Scale Integration have yielded incredibly precise manufacturing processes. They can create ever smaller yet highly consistent chips in which every transistor works exactly as intended. Thanks to them, our everyday electronics operate like clockwork despite the frailty of their overall architecture. However, even perfect manufacturing cannot prevent environmental damage. Computers exposed to the high energy radiation of outer space, or the hellish temperature and pressure of deep mines, must be hardened and duplicated to mitigate the risk of failure. These workarounds are costly and increase the complexity and power consumption of embedded computers in extreme environments.

Neuromorphic computing architectures, including Loihi, ROLLs, and FPGAs, mimic the organization of the brain with many neuron-like computing units. Even though these designs still feature several Single Point of Failure, they are a first step towards computing architectures that could sustain considerable damage while still working.

Outcomes:

During this thesis, the candidate will develop novel Neuromorphic hardware designs and algorithms and study their resilience to environmental damage. This will include the following steps:

• Design situational awareness algorithms for embedded computers in extreme environments (monitoring, navigation, mapping...) using Spiking Neural Networks and show that they can be implemented on Neuromorphic hardware.
• Evaluate the role of Spike-Timing Dependent Plasticity and other run-time local learning mechanisms in Spiking Neural Networks to compensate for the loss of computational units.
• Assess existing Neuromorphic hardware in terms of resilience to damage, and design optimized architectures with fewer points of failure.