On IoT Edge Devices: Manifold Unsupervised Learning for SoM Platforms

As developments in IoT edge devices become more on-demand, the technological advancements of these devices have been drastically increased. Accordingly, it became prominent to focus on integrating the AI processing on-board. Indeed, several modern platforms adopted the AI dedicated unit within their processor architectures (e.g. Xilinx SoM). Although these new architectures enabling the processing locally, the problem of sending the processed local data to the edge servers and cloud for further processing and decision making is still open. Precisely, sending a large amount of data out of the edge device will still consume a large amount of power and bandwidth. Herein, a system is proposed based on the modern SoM for edge devices to enable AI locally. The implementation is evaluated for an application connecting several cameras to the module, where the maximum CPU utilization reported is 77.6% and the minimum is 5.2%. Additionally, the Hyperalighnment algorithm is proposed to reduce the dimensionality of the locally processed data before sending it to the upper system levels.