【大咖观点】TelecomGPT: 电信领域的专属大模型

Large language models (LLMs) are expected to completely change the design paradigm of future wireless networks. However, the mainstream LLMs are generally lack of professional knowledge in the telecommunications field. To deal with this issue, a research team from Technology Innovation Institute (TII) and Khalifa University proposed for the first time a novel framework for applying general LLMs to the telecommunications field, called TelecomGPT. 

大语言模型（LLM）有望彻底改变未来无线网络的设计范式。然而，主流的LLM通常缺乏电信领域的专业知识。为了解决这一问题，来自阿布扎比技术创新研究所和哈利法大学的研究团队首次提出了一种将通用LLMs应用于电信领域的设计框架，称作TelecomGPT。

What are the unique features of TelecomGPT? How is it built? What are the potential applications of TelecomGPT? In this podcast, Prof. Merouane Debbah from Khalifa University, who is also the General Chair of the IEEE ComSoC GenAINet ETI, answered these questions, and showcased the potentials of the world’s first telecom-specific LLM: TelecomGPT.

The main points of this podcast are summarized below.

访谈的主要观点如下：

From general-purpose AI to telecom-specific AI, why?

Merouane Debbah: Telecommunication network is a super complex world, and general purpose AI is unable to understand the telecom language. For example, general purpose AI may have a basic knowledge about 5G, but it does not really understand how 5G works.

从通用AI到电信领域专属AI，为何需要这样的范式转变?

Merouane Debbah: 电信网络是一个超级复杂的世界，而通用AI无法理解电信语言。例如，通用的AI可能具有关于5G的基本知识，但是它并不真正理解5G是如何运行的。

How to build telecom-specific AI?

Merouane Debbah: First, we need massive dataset, called Open Telecom, which consists of standard documents, research papers, and even codes used in telecommunications. In this way, a dedicated library of telecommunication filed is established. However, it is not enough to only have this information. The key is how AI processes and learns from the information. The secret in building TelecomGPT is the three-stage training process. 

• Stage 1: Continuous Pre-training: Familiar AI with fundamental concepts of telecommunications knowledge. 

• Stage 2: Instruction Tuning: Learn to apply the knowledge in the telecom context. 

• Stage 3: Alignment Tuning: Refine the AI model to make sure that the outputs are accurate and relevant to the specific needs in the telecom field.

Merouane Debbah: 首先我们需要构建一个巨大的数据集，称作Open Telecom。这一数据集由大量的标准文档、研究论文、以及通信领域的代码构成。通过这种方式，将建立一个电信领域的专属数据库。然而，仅仅拥有信息是不够的，关键在于AI如何处理信息以及从信息中学习。构建TelecomGPT的奥秘在于如下的三阶段训练过程：

• 阶段1: 持续预训练——用于使AI熟悉电信领域的基础知识和基本概念； 

• 阶段2: 指令微调——用于使AI学会如何将具备的知识应用于电信领域上下文；

• 阶段3: 对齐微调——对AI模型进一步调优，从而确保其输出是精确的，并且对电信领域特定的任务和需求来说是有意义的。

Merouane Debbah: 我们开发了若干基线用于严格的性能评估。其中一个基线是数学建模。该基线用于测试TelecomGPT是否能够把对于网络的描述转化成数学方程。评估结果表明，经过微调的TelecomGPT在测试集上显著优于现有的大语言模型，例如GPT-4，尽管TelecomGPT的参数规模更小。其他的基线还包括开放问题问答和代码生成。 

What are the applications of TelecomGPT?

Merouane Debbah: 1) By combining the knowledge of the AI models with the analysis of actual radio signals, TelecomGPT can be used to interpret radio signals, analyze the signal strength, identify the interference, and maybe even decode the data in real-time, etc. 2) TelecomGPT is able to identify the subtle abnormal situations that people may miss. Therefore, TelecomGPT may be used to minimize network outage and optimize network performance. 3) TelecomGPT can be used as a tool for code generation. It can analyze the existing code, look for errors, and even generate new codes.

TelecomGPT未来会有怎样的应用?

Merouane Debbah: 1）通过将AI模型的知识与真实的无线信号分析相结合，TelecomGPT将能够用来解释无线信号，分析信号强度，识别干扰，甚至是实现实时的数据解码；2）TelecomGPT能够识别人类可能会忽略掉的微妙的异常，从而将能够用来最小化网络中断、优化网络性能；3）TelecomGPT可以被用作代码生成的工具。它可以被用来分析现有的代码，查找代码的错误，以及生成新的代码。

What are the impact on the future AI research?

Merouane Debbah: Building highly specialized AI could revolutionize other fields as well. For example, healthcare, finance, etc. In future, we may have a network of specialized AIs instead of a single super powerful AI.

TelecomGPT的出现对未来AI研究将会产生怎样的影响?

Merouane Debbah: 构建高度专业化的AI也将会对其他领域带来变革，例如：医疗、财经，等等。未来我们需要的或许并不是一个超级的AI，而是多个专属的AI，它们构成一个网络，相互协同工作。