Unequal loss protection for collaborative intelligence

This thesis explores combining Unequal Loss Protection (ULP) and Forward Error Correction (FEC) techniques in Collaborative Intelligence (CI). The focus is on reliably transmitting deep-feature tensor data in edge-cloud systems. The first part of the thesis tries to highlight a lack of detailed research on ULP within CI, despite its success in other contexts such as image and video transmission. As data integrity is crucial for CI systems, this work suggests a new way to make data transmission more robust against packet loss without needing more data to be sent. This is done by replacing less important data packets with FEC codes. These packets are chosen through a proxy model that mimics the modified Grad-CAM technique. The research also looks into how independent and identically distributed (iid) packet loss affects the accuracy of transmitting deep-feature tensor data. It shows that the new ULP method works well through extensive evaluations. Additionally, the thesis explores the integration of ULP and error concealment methods, focusing not only on the importance of packets but also on their reconstructability. The goal of this approach is to enhance the effectiveness of ULP by strategically selecting which packets to protect based on both how important they are and how easily they can be reconstructed if lost. This thesis helps advance CI by introducing a new method for using ULP in deep feature tensor transmission in edge-cloud systems.