Fundamental vocabulary in artificial intelligence and machine learning
As recognized by the AIQI Consortium, the ISO/IEC standards 22989 and 29119-11 are of high importance, because they contain the fundamental terminology relevant in conformity assessment and legal metrology:
- ISO/IEC 22989 “Information technology — Artificial intelligence — Artificial intelligence concepts and terminology”
- ISO/IEC 29119-11 “Technical specification describing AI/ML assessment metrics applicable to AI system life cycle stages”
AI and AI system
Artificial intelligence is the capability of an engineered system to acquire, process, and apply knowledge and skills (ISO/IEC 22989:2022)
In this sense, an artificial intelligence (AI) system is defined as an engineered system that utilizes AI technologies.
Machine learning is considered the foundational subset of AI. It utilizes computational algorithms that improve their performance through the ingestion of data rather than hard-coded programming.
The definition of “AI system” is of importance for the interpretation of regulatory requirements. For instance, in OIML Bulletin LXVI(3) 20250306, the author states that under the EU AI Act smart meters have to be considered to fall into the high-risk class when they utilize AI technologies, because they belong to critical infrastructure. With software separation, though, the AI part can be separated from the rest of the system, making it no longer an AI system in the strict sense.
Flavours of AI and ML
An artificial neural network is defined as a network composed of one or more layers of mathematically constructed neurons connected by weighted links with weights adjusted through the ingestion of data, i.e., in the training process. (ISO/IEC 22989:2022). There are many different categories of such neural networks, such as feed-forward neural networks, recurrent neural networks, and long short-term memory networks. For any neural network, input data is passed into an input layer and processed through the hidden layers up to the output layer.
Deep learning is basically a special subset of machine learning that uses neural networks possessing many hidden layers. A special category of deep learning is based on deep convolutional neural networks, which utilize mathematical convolution operations in their layers, making them particularly powerful for geometric and structural features in the input data.
Generative adversarial networks (GAN) combine two directly competing neural networks. A generator network synthesizes artificial data samples, while a discriminator network aims to distinguish synthetic data samples from authentic ones. Due to the combination of both networks, the generator’s training improves in synthesizing realistic data, i.e., synthetic data that is very close to the real-world. GANs are therefore often utilized to generate training data for other AI and ML.
An advanced deep learning system is a large language model (LLM). Its fundamental building block is the Transformer architecture that is trained on a huge amount of unstructured text data to learn intricate patterns, syntax, and semantics.
Agentic AI combines several independent AI models instead of relying on a single AI. The agentic AI distributes the tasks related to the input data among specialized, cooperating agents that work together via an orchestration layer.
The general challenge with more complex artificial intelligence models is their inherent opacity. The weights and biases of a deep learning model are considered type-specific parameters, making them relevant for conformity assessment (OIML D31). However, knowing these parameters doesn’t result in the capability to assess the model’s performance in complex scenarios. This “black-box” dilemma for conformity assessment and legal metrology is a challenge for any regulatory framework.
Generative adversarial networks and Large Language Models are examples of AI models that can be utilized in legal metrology workflows. For instance, Litzinger et al. (DOI 10.1016/j.measen.2024.101792) utilize an LLM to support analysing documentation as part of conformity assessment.
According to OIML D 31:2023, a dynamic module of legally relevant software is a module whose functional behaviour depends on predefined device-specific parameters that may change over time during use.
It is worth noting that such dynamic module may incorporate or utilize machine learning or artificial intelligence characteristics and processes.
