This thesis investigated the radiation tolerance of neuromorphic architectures. This was triggered by the findings from the existing literature that neuromorphic architectures have good radiation tolerance. To improve on this radiation tolerance, we introduced online learning, an Spike-Dependent Synaptic Plasticity (SDSP) learning mechanism that runs unsupervised on the system and has the task of counteracting radiation influences.
To validate our approach, we needed to expose our architecture to radiation. To make a test target for this, we needed an FPGA, for which we selected a flash-based FPGA. The device we picked was the PolarFire® SoC FPGA Icicle Kit. We also had to choose an open-source neuromorphic architecture to place on that FPGA; for this, we did a design space exploration and selected the Online-learning Digital spiking Neuromorphic processor (ODIN) architecture due to its learning capabilities and its relatively small size. We then created monitoring logic around the neuromorphic implementation and created a test setup. We also had to train the neuromorphic architecture and configure the online learning parameters. We selected the MNIST dataset for training. The task of the neuromorphic architecture was to classify those images. The initial model training was done with Python, the resulting weights of this training were put on the FPGA and that gave us the starting point for radiation testing. We created a UART connection to the FPGA to configure the architecture, set up the weights, and also to send the images and retrieve the classification output.
With our neuromorphic architecture setup on the FPGA device, we were able to test the architecture against radiation. We started with radiation beam testing at the ChipIR testing facility at Appleton Laboratory in the UK. This radiation beam allowed us to collect data about the radiation resistance of our system. We expanded the data collection by running a simulation based on the measured fault rate to get more information on radiation effects over a longer time period.
The ODIN architecture showed good radiation resistance without additional learning, but with added learning, the accuracy in the shorter runs decreased instead of increased. We ran additional simulations and found that over a longer period of time learning will have positive effects. We ran simulations until we had a completely failed system without learning. We did not see complete system failure with learning enabled, so this means that learning increases the time it will take for the system to completely fail by at least 175%. Therefore, while learning does not significantly contribute to preserving the system’s initial high level of accuracy against radiation effects, it is highly effective in substantially delaying the complete failure of the system.