端侧 AI：向边缘计算的智能转型（双语阅读）

本文首发于微信公众号：On-Device AI，欢迎关注，在这里，我们聚焦于设备上生成式人工智能的最新进展、技术深度解析、应用案例分享及未来趋势展望。

AI has been evolving from computers to smartphones, with each leap bringing us closer to smarter, more efficient technology. “On-device AI” is behind all this evolution. But what exactly is on-device AI, and why is it becoming so important?

人工智能已经从计算机发展到智能手机，每一次飞跃都让我们拥有更智能、更高效的技术。端侧 AI 是这一切进化的幕后推手。但究竟什么是端侧 AI 呢，它为何变得如此重要呢？

理解端侧 AI

On-device AI refers to the capability of edge devices like smartphones, smartwatches, and home assistants to perform artificial intelligence tasks directly on the device itself, without constantly needing to connect to the cloud or external servers. This means these devices can learn from data, make decisions, and take actions right where they are, using their built-in AI chips and software. While this might seem ordinary, it is actually quite significant, as there are several preliminary considerations to truly appreciate the value of on-device AI.

端侧 AI 指的是能够在像智能手机、智能手表和家庭助手这样的边缘设备上直接执行人工智能任务的能力，它不需要不断地连接到云端或外部服务器。这意味着这些设备可以使用内置的人工智能芯片和软件，直接从本地数据中学习、做出决策并采取行动。虽然这看起来很普通，但实际上意义重大，我们可以通过几个简单的论述来理解端侧 AI 的价值。

软件的部署模式

There are generally two conventional architectures for software deployment: cloud-based and on-premise. In a cloud-based setup, a server executes and controls everything, with clients simply sending requests to the server. Since all computations happen on the server, clients do not need to be computationally powerful. On the other hand, on-premise software runs directly on the client devices. This means that the software and its entire operational framework are housed within the device itself, which requires much more computational resources from the client.

软件部署通常有两种传统架构模式：基于云和本地部署。在基于云的部署模式中，服务器执行并控制一切，客户端仅向服务器发送请求。由于所有的计算都发生在服务器上，因此客户端不需要有很强的计算能力。另一方面，本地部署的软件直接在客户端设备上运行。这意味着软件及其整个操作框架都包含在设备内，这需要客户端拥有更多的计算资源。

On-device AI essentially bridges these two models by embedding AI capabilities directly into edge devices. It enables devices to perform immediate data processing without the need for external inputs, balancing the need for powerful local processing with the benefits of autonomy and instantaneity. This development not only mitigates the limitations associated with each traditional model—such as dependency on network connectivity and high latency in cloud architectures, and the high resource demands in on-premise setups—but also expands the potential applications of AI in everyday devices.

端侧 AI 本质上是通过将人工智能的能力直接嵌入到边缘设备中来桥接这两种模式。它使设备能够在不依赖外部输入的情况下立即进行数据处理，从而在强大的本地处理需求与自主性和即时性的优势之间取得平衡。这一发展不仅缓解了与传统模型相关的局限性——如对网络连接的依赖、云部署模式中的高延迟以及本地部署模式对机器资源的高性能要求，而且还扩展了 AI 在日常设备中的潜在应用。

基础技术

On-device AI is enabled by several key advancements:

1. Advanced Hardware: As Moore’s Law predicts, hardware is becoming more powerful and more affordable. High-end devices now include specialized AI chips tailored for efficient AI operations. These chips facilitate rapid AI task processing with minimal power use, making on-device AI a reality.
2. Optimized Software: Edge devices are more powerful than ever but still lag behind server capabilities. Despite this, AI algorithms have been effectively optimized for on-device execution. Techniques such as model pruning, which reduces the model size while maintaining performance, and quantization, which conserves memory by reducing the precision of calculations, enable these devices to run complex AI models efficiently.
3. Federated Learning: This machine learning framework trains a global model across various decentralized devices that hold local data samples, without exchanging raw data. This enhances data privacy and security and uses collective learning from multiple devices to improve model accuracy, all while avoiding heavy data transmission overheads.

端侧 AI 得益于几项关键的先进技术：

1. 强大的硬件：正如摩尔定律所预测的那样，硬件正变得越来越强大，价格也越来越实惠。高端设备现在内置了为高效 AI 操作量身定制的专用 AI 芯片。这些芯片以最小的功耗实现了快速的人工智能任务处理，使端侧 AI 成为现实。
2. 更优的软件：边缘设备比以往任何时候都更强大，但仍然落后于服务器功能。尽管如此，人工智能算法已经针对设备上执行进行了有效的优化。模型剪枝（在保持性能的同时减小模型大小）和量化（通过降低计算精度来节省内存）等技术使这些设备能够有效地运行复杂的 AI 模型。
3. 联邦学习：这个机器学习框架在各种分散的设备上训练一个全局模型，这些设备保存本地数据样本，而无需交换原始数据。这增强了数据隐私和安全性，并使用来自多个设备的集体学习来提高模型的准确性，同时避免了大量的数据传输开销。

These technologies empower devices to autonomously perform complex tasks such as voice recognition, language translation, and image processing. By using on-device AI, developers can create applications that are more responsive, efficient, and privacy-conscious, significantly expanding what smart devices can accomplish in everyday life.

这些技术使设备能够自主执行复杂的任务，如语音识别、语言翻译和图像处理。通过使用端侧 AI，开发人员可以创建响应更快、更高效、更注重隐私的应用程序，从而显著扩展智能设备在日常生活中可以完成的任务。

原文地址：https://azoo.ai/blogs/on-device-ai-the-bright-shifting-towards-the-edge