Information technology — Artificial intelligence — Artificial intelligence concepts and terminology
This international standard establishes terminology and describes concepts for artificial intelligence. It details the AI system life cycle, functional frameworks, and the broader ecosystem. The document provides standardised definitions for machine learning, neural networks, and trustworthiness characteristics to support communication between global stakeholders and assist in system classification.
Please login or join for free to read more.
OVERVIEW
Introduction
Advancements in computing capacity, reduction in costs, and the availability of vast datasets have enabled practical applications of artificial intelligence (AI), making it a critical branch of information technology. AI is an interdisciplinary field involving computer science, data science, mathematics, and social sciences. This document provides standardised concepts and terminology to ensure AI technology is better understood by stakeholders, including experts and non-practitioners. It defines AI as a technical and scientific field devoted to engineered systems that generate outputs such as content, forecasts, recommendations, or decisions for human-defined objectives.
Scope
This document establishes terminology and describes concepts within the field of AI. It is intended to support communications among diverse parties and is applicable to all types of organisations, including commercial enterprises, government agencies, and not-for-profit organisations.
Terms and definitions
The report categorises essential terms into several clusters: AI-related terms, data-related terms, machine learning (ML), neural networks, trustworthiness, natural language processing (NLP), and computer vision. Key definitions include an AI system as an engineered system designed to operate with varying levels of automation. Machine learning is defined as a process of optimising model parameters through computational techniques so that the model’s behaviour reflects data or experience.
AI concepts
AI systems are characterised by features such as interactivity, contextuality, oversight, and adaptivity. A fundamental distinction is made between narrow AI, which is focused on defined tasks to address specific problems, and general AI, which addresses a broad range of tasks with satisfactory performance. Currently, AI systems are considered narrow; it remains unknown if general AI will be technically feasible in the future. The report also details the agent paradigm, where an AI agent interacts with its environment via sensors and actuators to achieve goals.
Data
Data is central to AI systems, particularly machine learning. The document outlines the data life cycle processes, including acquisition, exploratory data analysis, annotation, preparation, filtering, normalisation, and quality checking. It distinguishes between production data (processed during operation), test data (used for performance evaluation), and training data (used to extract the model). Addressing data drift and concept drift is highlighted as a necessity for maintaining model accuracy over time.
Machine learning concepts
The report describes various ML approaches: supervised learning (using labelled data), unsupervised learning (using unlabelled data), and reinforcement learning (learning through trial and error to maximise rewards). It introduces the concept of “catastrophic forgetting,” a limitation where a model overwrites previously acquired knowledge when introduced to new training data. Continuous learning is presented as a special case of retraining that occurs incrementally while the system is in production.
Trustworthiness
Trustworthiness in AI refers to the ability to meet stakeholder expectations in a verifiable way. Key characteristics include robustness (the ability to maintain performance levels under any circumstances), reliability (consistent intended behaviour), and resilience (the ability to recover quickly from incidents). It also encompasses explainability, which allows humans to understand the factors influencing a decision, and transparency, ensuring appropriate information about the system is available to stakeholders. The document notes that minimising unwanted bias is a challenging but essential goal for fair system results.
AI system life cycle
The AI system life cycle model describes evolution from inception through to retirement. It is often more iterative than traditional software life cycles. Stages include inception (defining objectives and requirements), design and development, verification and validation (testing functionality and capabilities), deployment, operation and monitoring, and re-evaluation. Continuous validation is required for systems using continuous learning to ensure they operate correctly as they adapt to new data.
AI system functional overview
Functionally, AI systems contain a model used to produce predictions. These predictions are primary outputs that are then used to make recommendations, decisions, and actions. The report emphasises that AI outputs are error-prone and have a probability of being correct rather than being absolutely true. Human oversight is typically present during training and validation to ensure systems are developed and used as intended.
AI ecosystem
The AI ecosystem is represented as functional layers, from hardware resource pools (including CPUs, GPUs, and ASICs) to big data sources and specific AI functions like reasoning. Application-specific integrated circuits (ASICs) are highlighted for providing high computational capability with lower energy consumption, particularly in mobile and IoT devices.
Fields of AI
Major fields include computer vision, which involves acquiring and interpreting visual data, and NLP, which enables computers to analyse and generate human language. NLP components include natural language understanding (NLU) and natural language generation (NLG), supporting applications like speech recognition and machine translation.
Applications of AI systems
Practical applications discussed include fraud detection (identifying fraudulent charges), automated vehicles (utilising computer vision for object avoidance), and predictive maintenance (detecting potential failures in infrastructure or machinery based on observed behaviour).