On Metric-Driven Development of Embedded Neuromorphic AI

Neuromorphic computing presents a promising approach to embedded artificial intelligence (AI) by deploying spiking neural networks (SNNs) to specialized accelerators. Such neuromorphic accelerators embrace the unique properties of SNNs to enable applications with exceptionally little memory footprint, energy demand, and inference time. By increasing their naturally high temporal sparsity and using popular approaches to improve their spatial sparsity, this can be taken to the extreme. However, improving hardware-related quantities always comes with the tradeoff of degrading accuracy. While available literature barely tolerates any loss in accuracy, in this work, we argue that this might be acceptable for optimal performance of embedded neuromorphic systems, depending on the use case. We propose an approach centered around the maximization of metrics that contain all application-specific requirements and discuss the possibilities of designing such a metric. Finally, we consider three scenarios and demonstrate the execution of our approach based on the neuromorphic DVS128 Gesture dataset. The resulting networks’ properties are tailored to the respective applications’ needs and highlight the effectiveness of our approach. Our code is available at https://github.com/JannKrausse/MetricDrivenNeuromorphics.