Standards Development

Scope

This document provides guidance on developing lightweight model ling for AI systems. It describes a framework including common concepts, components, processes, and assessments of AI lightweight modelling. It does not specify any detailed models. 

Purpose

This document describes guidance for AI lightweight modelling, which is the process of developing and deploying optimized AI models that maintain satisfactory accuracy and functionality while optimizing for resource efficiency and computational performance. Maintaining the performance the same as without reducing AI system even though an AI model size is reduced is a very challenging issue.

To ensure high performance in AI systems, large-scale data and various AI technologies are essential. However, the industrial sectors applying AI are diverse not only in terms of hardware and software environments but also require operation in both large-scale environments like the cloud and small - scale devices such as IoT, sensors, and more. For these reasons, many AI systems often opt for methods that reduce model size, data volume, parameters, and so on. However, this can sometimes lead to the underperformance in AI systems, especially in constrained scenarios like device limitations. Additionally, even with fully equipped hardware, the large computational demands often exist in an inefficient state without optimization. Therefore, this standard document focuses on providing optimized AI lightweight guidelines that allow for lightweight according to industrial application areas while ensuring the high performance of AI systems. It also aims to present the benefits and perspectives achievable through lightweight.

This standard is applicable not only to small -scale but also to large-scale AI systems, optimizing their performance in many industrial applications. However, even though AI lightweight is a challenging and useful endeavour, there are currently no guidelines on what items should be lightweight, what constitutes optimal lightweight, and how to evaluate lightweight. This document offers standard guidance on the AI lightweight framework, encompassing metrics for AI lightweight. The goal of AI lightweight is not merely to reduce, but to actually high performance and optimization of AI systems. This is accomplished by creating models that are efficient, consume fewer resources, and are well-suited for deployment in diverse computing environments. Overall, AI lightweight plays a crucial role in realizing the full potential of AI across different domains and platforms. The ways to achieve "AI lightweight" involve minimizing hardware requirements, such as reducing battery and memory usage and simplifying architecture. Additionally, it includes optimizing resource demands, including computational time, model size, parameter count, and data transfer, especially in small-scale, cloud, edge, IoT, and sensor environments. At the same time, the goal is to attain high performance characterized by low latency, cost -effectiveness, energy efficiency, and superior metrics like accuracy, speed, adaptability, and throughput.

This document aims to provide comprehensive guidance and standards for AI lightweight modelling, encompassing the entire process from thorough requirement analysis, including considerations of resource and infrastructural needs, to the final stage of model assessment. The assessment involves metrics for the performance of model optimization, such as accuracy, adaptability, efficiency, speed, throughput, cost, energy efficiency, latency, and more, as illustrated in the optimized AI lightweight modelling framework shown in Figure 1 (See Annex - Figure 1). The optimization process encompasses model preparation, involving feature engineering, model design, incorporating model selection, and model training, which includes algorithm selection.

This document focuses on the guidance on AI lightweight modelling. ISO/IEC 23053:2022 primarily addresses AI systems for machine learning (ML), but the concept of AI lightweight modelling is not explicitly defined in ISO/IEC 23053:2022, despite mentioning optimization. While ISO/IEC FDIS 5338:2023 covers the AI system life cycle process and ISO/IEC 8183:2023 pertains to the data life cycle framework derived from ISO/IEC 23053:2022.

The ToRs are as follows.

— The concepts and terminology for AI lightweight modelling

— AI lightweight modelling framework:

- Classification of lightweight

- AI lightweight methods and optimization process for model preparation (feature engineering), model design (model selection) and model training (algorithm selection)

— The assessment of AI lightweight modelling:

- Assessment metrics for model assessment

- Assessment measuring methods

This standard aims to develop AI lightweight modelling that covers an overview, key components of AI lightweight modelling, relevant concepts and terminology, and the overall processes and assessments involved in the framework.

Justification

In fact, there are increasing many parameters and data to be processed when modelling AI systems. As time goes by, the computational requirements of AI are rapidly increasing. To maintain accuracy, AI models require substantial computation and training time, a demand that's outpacing the enhancements from GPU technology. To achieve high performance, AI systems often connect with other devices or networks, but as the size of AI connection increases, it becomes crucial to address the challenge of reducing computation time. AI lightweight modelling, with fewer parameters yet maintaining comparable performance, can help tackle this issue by speeding up computations and enhancing efficiency in these large-scale networks. AI lightweight can be addressed through hardware methods, but even with the same hardware specifications, efforts are needed to address it through software methods.

The advantages of standard AI lightweight modelling encompass the enhancement of AI modelling by establishing fundamental concepts, the reduction of the development costs of AI systems, and its impact on the broader AI ecosystem, industrial AI domains, and other AI inference technologies, etc. AI systems often impose significant demands on computational resources, resulting in elevated costs and infrastructure requirements. The strategic adoption of AI lightweight modelling seeks to optimize resource utilization, diminish computational complexity, and, consequently, achieve substantial cost savings in hardware, energy consumption, and overall infrastructure.

The justification for embracing AI lightweight modelling is rooted in its ability to effectively tackle challenges arising from resource limitations, cost constraints, and the imperative for efficient AI system inference. Through these optimizations, AI lightweight modelling not only addresses these challenges but also enhances the effectiveness and accessibility of AI solutions.

AI lightweight models play a pivotal role in enhancing computational efficiency, enabling real -time processing, and facilitating seamless deployment of AI on edge devices. By curtailing computational and energy demands, these models significantly bolster performance, particularly in time -sensitive applications such as autonomous vehicles, robotics, and healthcare monitoring. Moreover, their contribution to prolonging battery life and minimizing environmental impacts underscores their sustainability.

The adaptability of lightweight models to operate efficiently on devices with constrained resources not only makes AI technology more accessible but also greatly enhances the user experience in interactive applications, reducing latency. By adopting AI lightweight modelling practices, organizations can realize the dual benefits of efficient and effective AI systems, aligning with the evolving landscape of technology and user expectations.