Application of Neuromorphic Computing to Real World Scenarios

Spiking neural networks (SNNs) aim to mimic the human brain in the way that they process data. Neurons connected via synapses fire, or spike, sending information to other neurons and producing an output. SNNs have been applied to a variety of problem spaces for tasks such as classification and control. They are able to process real-time data efficiently and are particularly useful for application to temporal data. This work investigates the use of hafnia-based ferroelectric devices as synapses for a pole-balancing control task in varying levels of difficulty. We found that the ferroelectric networks are able to achieve comparable performance to basic neuromorphic implementations, indicating that these devices are well-suited to neuromorphic computation and performance limitations can be overcome with additional optimization. The best performing read voltage was task-dependent but performance could be improved with additional training. Additionally, the inclusion of noise in the ferroelectric device data, which more accurately reflects the behavior of these devices in the real world, did not significantly impair performance and in some cases even produced better results. We also found that small networks with relatively low activity still achieved good performance. We used SNNs to process experimental minute ventilation data for mice of varying sex, buprenorphine dosage, obesity, and leptin statuses. We attempted to classify buprenorphine dosage, congenic line, and sex using the minute ventilation readings. We found that neuromorphic approaches were more accurate in general than standard machine learning approaches. Our results show that including the temporal component of the data significantly improved performance due to the temporal nature of SNNs.