TinyML Empowered Transfer Learning on the Edge
TinyML Empowered Transfer Learning on the Edge
复制标题
DOI：
10.1109/ojcoms.2024.3373177
复制
发表时间：
2024
影响因子：
7.9
通讯作者：
A. Hayajneh;Maryam Hafeez;Syed Ali Raza Zaidi;Des McLernon
中科院分区：
文献类型：
--
作者：
A. Hayajneh;Maryam Hafeez;Syed Ali Raza Zaidi;Des McLernon
微型机器学习 (TinyML) 是一种很有前途的方法，可以在资源有限的低功耗物联网 (IoT) 边缘设备上实现依赖人类活动识别 (HAR) 的智能应用程序。然而，由于计算资源限制以及需要针对独特用例进行定制，为这些设备设计高效的 TinyML 模型仍然具有挑战性。为了解决这个问题，我们提出了一种新方法，在边缘微控制器单元 (MCU) 上利用迁移学习 (TL) 技术来加速 TinyML 开发。我们的策略包括在大规模和多样化的数据集上预训练广义机器学习模型，并使用 TL 在设备上针对特定应用程序对其进行微调。我们通过实验卷积神经网络、长短期记忆 TL (CNN-LSTM-TL) 模型架构和用于降维的 t 分布随机邻域嵌入 (t-SNE) 等可视化技术，证明了我们的 HAR 方法的有效性。为了进一步验证我们模型的熟练程度和适应性，我们使用两个不同的数据集：MotionSense 和 UCI 进行了广泛的测试。这种双数据集方法使我们能够评估模型在不同数据域中的稳健性，展示其在各种 HAR 场景中的多功能性和有效性。我们的结果显示模型准确性显着，训练时间缩短，同时保持高推理率和低 MCU 内存占用。我们还提供有关在边缘 MCU 上实施 TL 的最佳实践的见解，并评估分类性能指标，例如准确度、精确度、召回率、F1 分数和分类交叉熵损失。我们的工作为通过 TL 框架在不同应用领域和边缘物联网设备类型上更快、更高效地部署 TinyML 奠定了坚实的基础。
Tiny machine learning (TinyML) is a promising approach to enable intelligent applications relying on Human Activity Recognition (HAR) on resource-limited and low-power Internet of Things (IoT) edge devices. However, designing efficient TinyML models for these devices remains challenging due to computational resource constraints and the need for customisation to unique use cases. To address this, we propose a novel approach that utilises transfer learning (TL) techniques on edge micro-controller units (MCUs) to accelerate TinyML development. Our strategy involves pre-training generalised ML models on large-scale and varied datasets and fine-tuning them on-device for specific applications using TL. We demonstrate the effectiveness of our approach for HAR by experimenting with a convolutional neural network, long short-term memory TL (CNN-LSTM-TL) model architecture and visualisation techniques like t-distributed stochastic neighbour embedding (t-SNE) for dimensionality reduction. To further validate our model’s proficiency and adaptability, we conducted extensive testing using two distinct datasets: MotionSense and UCI. This dual-dataset approach allowed us to assess the robustness of our model across different data domains, showcasing its versatility and effectiveness in various HAR scenarios. Our results show significant model accuracy and reduced training time while maintaining high inference rates and low MCU memory footprint. We also provide insights into best practices for implementing TL on edge MCUs and evaluate classification performance metrics such as accuracy, precision, recall, F1 score, and categorical cross-entropy loss. Our work lays a solid foundation for faster and more efficient TinyML deployments through TL framework across different application domains and types of edge IoT devices.