Standard

SAE J3321

Publisert
For bestilling og priser av dette produktet ta kontakt med salg@standard.no

Omfang

AI/ML Systems This document intends to focus on AI/ML systems deployed on board and/or off board the vehicles, including the methods and tools used anywhere in the life cycle of automotive system development. This includes, for instance, AI/ML systems used in the identification of parameters, calibration of classical or neural network-based control algorithms, system level modeling and validation, data process automation pipeline such as automated labeling and annotation, data quality checks, etc. Furthermore, with the continued success of Generative AI (GenAI) capabilities, it is becoming a common practice to leverage these capabilities beyond just generating synthetic data. Multiple artifacts such as code, testing scenarios, and benchmarks could be developed using GenAI tools, necessitating the additional verification of these “generative” development steps. The fundamental challenge with generative methods is grounding, especially if data, scenes, and tests are used for the development of AI-enabled products and features. AI and ML systems have been the subject of several national and international standards and guideline documents. A brief selection of these are given in Section 2. One, ISO PAS 8800:2024, focuses on safety over the entire life cycle of AI development and deployment. This document is intended to serve the following purpose: a. The document is an information report with no mandatory requirements. b. The focus of the document is only on V&V of the development life cycle of AI-based components or systems. c. The V&V methods are general in nature. The methods discussed include, but are not restricted to, safety requirements or properties. V&V of AI/ML-Based Systems V&V of AI/ML-based components and systems differs from that of conventional software and systems and requires special attention. Unlike conventional software, AI/ML-based systems often use very complex, high dimensional, and unstructured information sources as inputs (e.g., images, video, audio, natural language, etc.). Additionally, these systems often grapple with human-based concepts of cognition and reasoning, perform functions such as analysis or processing (e.g., detection of vehicles, pedestrians and animals, prediction of accidents and traffic jams, understanding speech and text), and perform analysis of sentiments and prediction of accidents, etc. More importantly, these systems can assist with decisions or increasingly take decisions autonomously (agentic AI). These systems often employ probabilistic algorithms or models (neural networks, graphs, Bayesian networks) and produce stochastic outputs with associated confidence levels. For such systems, changes in the model inputs that are outside of the model training/validation dataset may produce different results that are misaligned with the expectation. The degree of deviation is a function of the model capacity or expressivity and a function of distance of the distribution of the perturbed inputs relative to the training/validation data. Such out of distribution (OOD) of shifted inputs is often observed via real-world phenomenon or events that may be subjected to uncertainties and noise. The models themselves are constructed using a large but often incomplete sample representation of the real-world inputs which also can/will vary over time. The models contain parameters and rely on stochastic optimization methods to compute appropriates values for these parameters that optimize an objective function (e.g., loss function, likelihood function). In view of the above differences relative to traditional deterministic models, there is a strong need for reviewing existing V&V methods and developing new methods and/or adapting existing methods and best practices for a robust V&V of data-driven AI models. Quality and Dependability The overarching theme of this report is to provide detailed information on V&V methods for ensuring quality and enabling the development of reliable AI-enabled systems. There are several quality attributes expected by the end users and included either implicitly or explicitly in the requirements of a system. The quality requirements identify several quality attributes that the system is expected to have and, depending upon the context and application, some or all of these attributes that the system under test is expected to have. In the context of AI/ML applications, some of the quality attributes that are often emphasized are robustness, generalizability, and explainability, in addition to accuracy, correctness, and performance. For the sake of keeping the report succinct, it focusses less on the details of different quality attributes and more on their verification. We further restrict the report to the V&V of only a few selected quality attributes, which include functionality, safety, and robustness. We rely on some recent standards and reports on the broader topic of quality, reliability, and dependability published by the ISO/SAE/IEC/IEEE communities. These include three ISO publications: ISO/IEC 25010:2023, ISO/IEC TS 25058:2023, and ISO/IEC 25059:2023.

Dokumentinformasjon

  • Standard fra SAE_AC
  • Publisert:
  • Versjon: 0
  • Varetype: IS