Authors:

Peter Zakariya

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

Sensor networks are systems that collaboratively monitor an area, collecting and processing data in harsh environments with limited communication. Integrating machine learning with sensor networks enhances real-world applications such as remote wildlife monitoring or anomaly detection for security around a building. The challenges of a harsh environment motivate the need for robustness under restricted bandwidth. To understand the effects of different approaches in this setting, we simulated sensor networks with an image segmentation task, where each pixel is a node and the communication topology is created through Erdos-Rényi or spatial grid generation to represent different network structures. During each communication round, nodes are constrained to send one message to one neighbor. Nodes are able to aggregate partial information to create a representation of their surrounding environment; in our simulation, nodes create a corrupted version of the original image. U-Net models were leveraged to predict the pixel's label at a node's location, and as expected, accuracy improves over communication rounds. Sending only one pixel in a message is inefficient, motivating the need for using techniques to encode more than just raw information in a message. One technique being explored is wavelet transforms, a signal compression technique commonly used in image compression. Further strategies are explored to prioritize selecting neighbors who would gain the most information from the message to increase the information exchanged per communication round. This research advances the application of decentralized learning to real-world scenarios with limited communication between devices, where maximizing the information of each message is crucial. Keywords: Machine Learning; Deep Learning; Decentralized; Networks; Sensors

Source:

Purdue University / 2025

Topics:

No topics listed

Co-authors:

Peter Zakariya