Exploring Edge TPU for deep feed-forward neural networks

This paper explores the performance of Google's Edge TPU on feed forward neural networks. We consider as a hardware platform and explore different architectures deep network classifiers, which traditionally has been challenge to run resource constrained edge devices. Based use joint-time-frequency data representation, also known spectrogram, we trade-off between classification energy consumed for inference. The efficiency is compared with that widely-used embedded CPU ARM Cortex-A53. Our results quantify impact architectural specifications TPU's performance, guiding decisions optimal operating point, where it can provide high accuracy minimal consumption. Also, our evaluations highlight crossover in Cortex-A53, depending specifications. analysis, decision chart guide selection based model parameters context.