SHARED LEARNING AMONG DISTRIBUTED EDGE DEVICES USING CORAL EDGE TPU MACHINE LEARNING ENGINES

Internet of Things (IoT) edge devices have small amounts of memory and limited computational power. These resource-constrained devices consist of sensors that generate large amounts of data, making IoT edge devices attractive targets for machine learning models. To take advantage of machine learning models normally requires the data to be transported to a remote device with enough computational power to process these data. The transport of data to a remote node creates a delayed response and is dependent on data transport availability. Besides performance hits to machine learning models on IoT at the edge, any model training on IoT edge devices is nearly impossible. With the introduction of the Coral Tensor Processing Unit (TPU), real-time data processing through machine learning models on IoT edge devices is achievable. This research explores splitting a convolutional neural network (CNN) to expose an intermediate layer for fine-tune training. This study found that it is possible to extract an intermediate layer output from a CNN running on the TPU for fine-tune training on a Raspberry Pi v4 where the fine-tuning is done only on the upper layers of the model. This makes it possible to fine-tune train larger models on a resource restricted device. The model’s performance improved 6.7%, from 53.9% to 60.6%.Major, United States ArmyApproved for public release. Distribution is unlimited