...
- Need for High Quality Structured Data: Telecommunications networks are very different from other AI human-computer data sets, in that a large number of interactions between systems use structured data. However, due to the complexity of network systems and differences in vendor implementations, the degree of standardization of these structured data is currently very low, causing "information islands" that cannot be uniformly "interpreted" across the systems. There is a lack of "standard bridge" to establish correlation between them, and it cannot be used as effective input to incubate modern data-driven AI/ML applications.
- AI Trustworthiness: In order to meet carrier-level reliability requirements, network operation management needs to be strict, precise, and prudent. Although operators have introduced various automation methods in the process of building autonomous networks, organizations with real people are still responsible for ensuring the quality of communication and network services. In other words, AI is still assisting people and is not advanced enough to replace people. Not only because the loss caused by the AI algorithm itself cannot be defined as the responsible party (the developer or the user of the algorithm), but also because the deep learning models based on the AI/ML algorithms themselves are based on mathematical statistical characteristics, resulting in sometimes erratic behavior, meaning the credibility of the results can be difficult to determine.
- Uneconomic Uneconomical Margin Costs: According to the analysis and research in the Intelligent Networking, AI and Machine Learning While Paper , there are a large number of potential network AI technology application scenarios, but to independently build a customized data-driven AI/ML model for each specific scenario, is uneconomical and hence unsustainable, both for research and operations. Determining how to build an effective business model between basic computing power providers, general AI/ML capability providers and algorithm application consumers is an essential prerequisite for its effective application in the field.
- Unsupportable Research Models: Compared with traditional data-driven dedicated AI/ML models in specific scenarios, the research and operation of large language models have higher requirements for pre-training data scales, training computing power cluster scale, fine-tuning engineering and other resource requirements, energy consumption, management and maintenance, etc. It is difficult to determine if it is possible to build a shared research and development application, operation and maintenance model for the telecommunications industry so that it can become a stepping stone for operators to realize useful autonomous networks.
- Contextual data sets: Another hurtle that is often overlooked is the need for the networking data sets to be understood in context. What that means is that networks need to work with all the layers of the IT stack, including but not limited to:
- Applications: Making sure that customer applications perform as expected withe underlying network
- Security: More important than ever as attack vector expand and customers expect the networks to be protected
- Interoperability: The data sets must support transparent interoperability with other operators, cloud providers and other systems in the telecom ecosystem
- OSS/BSS Systems: The operational and business applications that support network services
...
In addition to general domain adaptation, many telcos will have the issue of multilingual datasets, where a mix of languages (typically English + something else) will exist in the data (syslogs, wikis, tickets, chat conversations etc.). While many options exist for both generative LLMs [11] and text embedding models [12], not many foundation models have seen enough non-English data in training, thus options in foundation model choice are somewhat restricted for operators working on non-English data. A solution to work around this issue is to use automated translation and language detection models on the data as a preprocessing step.
<sorry for the late addition, feel free to edit or delete Andrei Agapi Beth Cohen - added by Jason Hunt >
Beyond these approaches, an emerging technique of for pre-training foundation models on network data holds potential promise. In With this technique, network data is essentially turned into a language via pre-processing and tokenization, which can then be used for pre-training a new "network foundation model." [34] Initial research has successfully demonstrated this technique approach on Domain Name Service (DNS) data [35] and geospatial data [36]. As this area of research matures, it could allow for general purpose network foundation foundational models that can be fine-tuned to answer a variety of questions around network data or configurations without having to train individual models for bespoke network management tasks. <end addition>
In conclusion, while foundation models and transfer learning have been shown to work very well on general human language when pre-training is done on large corpuses of human text (such as Wikipedia, or the Pile [29]), it remains an open question to be answered whether domain adaptation and downstream task adaptation work equally well on the kinds of domain-specific, semi-structured, mixed modality datasets we find in the telecom industry. To enable this, telecoms should focus on standardization and data governance efforts, such as standardized and unified data collection policies and developing high quality structured data across a common definition and understanding.
...