SemanticNN: Compressive and Error-Resilient Semantic Offloading for Extremely Weak Devices

TL;DR

SemanticNN tackles the challenge of running AI inference on extremely weak devices over unreliable wireless links by offloading compressed intermediate features and tolerating bit errors at the semantic level. The authors design a SQ-based encoder and BER-aware decoder, combined with Feature-augmentation Learning and XAI-based asymmetry compensation to preserve semantic fidelity under transmission errors. Experimental results on STM32 with multiple models and datasets show massive reduction in offloaded data (56.82-344.83x) while maintaining or surpassing accuracy under BERs, across image classification and object detection tasks. Real-world case studies in LoRa and indoor Wi-Fi confirm robustness to dynamic BERs, highlighting practical impact for edge intelligence in IoT.

Abstract

With the rapid growth of the Internet of Things (IoT), integrating artificial intelligence (AI) on extremely weak embedded devices has garnered significant attention, enabling improved real-time performance and enhanced data privacy. However, the resource limitations of such devices and unreliable network conditions necessitate error-resilient device-edge collaboration systems. Traditional approaches focus on bit-level transmission correctness, which can be inefficient under dynamic channel conditions. In contrast, we propose SemanticNN, a semantic codec that tolerates bit-level errors in pursuit of semantic-level correctness, enabling compressive and resilient collaborative inference offloading under strict computational and communication constraints. It incorporates a Bit Error Rate (BER)-aware decoder that adapts to dynamic channel conditions and a Soft Quantization (SQ)-based encoder to learn compact representations. Building on this architecture, we introduce Feature-augmentation Learning, a novel training strategy that enhances offloading efficiency. To address encoder-decoder capability mismatches from asymmetric resources, we propose XAI-based Asymmetry Compensation to enhance decoding semantic fidelity. We conduct extensive experiments on STM32 using three models and six datasets across image classification and object detection tasks. Experimental results demonstrate that, under varying transmission error rates, SemanticNN significantly reduces feature transmission volume by 56.82-344.83x while maintaining superior inference accuracy.

Figures (18)

• Figure 1: SemanticNN (including encoder and decoder) accepts bit errors in wireless transmissions, pursuing semantic correctness with extremely feature compression in split computing. The original task NN will be frozen, and only the SemanticNN will be optimized.
• Figure 2: The detailed model structures of the SQ-based encoder and BER-aware decoder in SemanticNN.
• Figure 3: Balanced quantization with eight learning quantization centers and 3-bit compact coding.
• Figure 4: Feature-augmentation Learning first performs Denoising Autoencoder Training to initially fight against transmission errors. Subsequently, Task-Oriented Semantic-level Training is performed to extract and transmit semantic information on specific downstream tasks for better performance.
• Figure 5: Self-aware pixel-level feature importance analysis using XAI and partial feature offloading.