SyncEdge: Deadline-Driven Zero-shot Recognition with Dynamic Object Sharing on Multiple Edge Devices = SyncEdge: deadline에 따른 제로샷 인식을 위한 다중 엣지 장치에서의 동적 객체 공유

다국어 초록 (Multilingual Abstract)

In recent years, extensive research efforts have been dedicated to exploring zero-shot learning techniques aimed at facilitating the identification of unseen categories that were not included in the training data. One promising technique in enhancing accuracy is Context-Aware Zero-Shot Recognition. By incorporating local context information, including the relationship between multiple objects, this approach aims to provide a more comprehensive understanding of the data. However, when implementing this technique on edge devices within Internet of Things (IoT) and embedded systems, significant challenges such as critical deadline violation arise due to high computation requirements.
To address these challenges, we introduce SyncEdge, a novel runtime framework designed to enable On-device Context-Aware Zero-Shot Recognition. SyncEdge leverages the concept of frame partitioning, which enables the distribution of computational tasks across multiple edge devices. By effectively sharing object information among these devices, SyncEdge optimizes the overall recognition process without time violations. To evaluate the effectiveness of SyncEdge, we conducted experiments in two distinct settings. Firstly, we utilized two NVIDIA Jetson Xavier AGX boards within a homogeneous environment. Secondly, we employed one AGX board and one NX board in a heterogeneous environment. In both cases, SyncEdge demonstrated tangible improvements in accuracy by up to 1.89%.

목차 (Table of Contents)

• Chapter 1. Introduction 1
• Chapter 2. Background 4
• 2.1 Zero-shot learning 4
• 2.2 Context-Aware Zero-shot Recognition 5
• Chapter 3. Motivation 6
• 3.1. Context-Aware Zero-shot Recognition on Single-Device 6
• 3.2. Relationship Inference 8
• Chapter 4. SyncEdge Architecture 9
• 4.1. Data pre-processing 10
• 4.2. Sharing Object information 10
• Chapter 5. Experimental Results 16
• 5.1. Experimental Set-up 16
• 5.2. Speed up of SyncEdge 17
• 5.3. Accuracy of SyncEdge 25
• Chapter 6. Conclusion 28
• References 29
• Abstract in Korean 31