A Neural Interface SoC for Smart Glasses with Low-Power Neural Commanding and Efficient LoRA-Enabled On-Chip Learning

📋 论文概要

该论文提出一款用于智能眼镜的65nm神经接口SoC，集成10通道AFE和片上CNN，实现实时EOG/EMG/EEG推理。采用LoRA微调的片上持续学习将可训练参数减少54-77%，并设计超低功耗最近特征神经命令路径以按需唤醒重计算，同时通过分块delta编码和流式流水线减少内存并扩展训练数据2.1倍。

💡 主要创新点

• 采用LoRA微调实现片上持续学习，将可训练参数削减54-77%，显著降低计算开销。
• 设计超低功耗最近特征神经命令路径，仅在需要用户命令时唤醒重计算，降低系统功耗。
• 提出分块delta编码和流式流水线，减少内存占用并扩展训练数据量2.1倍。

🏷 关键词

📄 原文摘要

*Equally Credited Authors (ECAs) 1 Abstract This work presents a 65nm ExG SoC for smart glasses, enabling low-power neural interaction. A 10-ch AFE and on-chip CNN deliver real-time EOG/EMG/EEG inference. Ondevice continual learning via LoRA fine-tuning cuts trainable parameters by 54 to 77%. An ultra-lower-power nearest-feature neural commanding path was implemented to issue user commands and wake up heavy computation only when needed. Chunked delta encoding and a streaming pipeline reduce memory and expand training data by 2.1×. Smart glasses powered by augmented reality (AR) and artificial intelligence (AI) are projected as a new market driver for wearable devices, offering a versatile platform for consumer applications [1,2]. However, current smart glasses still fall short of delivering effortless ergonomic human interactions. First, as shown in Fig. 36.2.1, the existing smart glasses interface with users through physical touchpads or voice commands, which require

👥 作者与机构

Zhiwei Zhong*1, He Yu*1, William McGarry1, Yijie Wei2, Jie Gu1

Northwestern University, Evanston, IL, 2Texas Instruments, Dallas, TX