|
Work item:
|
Y.AMLM-reqts
|
|
Subject/title:
|
Cloud computing - Functional requirements for adaptation of machine learning model
|
|
Status:
|
Under study
|
|
Approval process:
|
AAP
|
|
Type of work item:
|
Recommendation
|
|
Version:
|
New
|
|
Equivalent number:
|
-
|
|
Timing:
|
2027-Q1 (High priority)
|
|
Liaison:
|
ITU-T SG20, SG21, ITU-R SG6, ISO/IEC JTC 1/SC SC29, SC38, SC42, W3C, Khronous Group
|
|
Supporting members:
|
China Telecom, ETRI, Beijing University of Posts & Telecommunications, Zhejiang Lab
|
|
Summary:
|
With the rapid adoption of Machine Learning as a Service (MLaaS) in cloud computing, the focus of ML development has shifted from model creation to application development. MLaaS frameworks (described in ITU-T Y.3531) primarily support users in training ML models from scratch. They provide tools for data processing, model training, and deployment—either in cloud or on-premises. However, MLaaS does not specifically address the needs of ML model usage phase, where trained models need to be adapted, delivered and integrated for real-world applications.
These users face critical pain points:
High barriers: Training models demands ML expertise and substantial time and resources.
Resource waste: Re-training ML models (e.g., image classification, text generation) is inefficient.
Adaptation complexity: Adapting trained ML models to specific scenarios involves compatibility issues, cross-domain customization hurdles, and task-specific adaptation obstacle.
In contrast, adaptation of ML model begins where MLaaS ends, after a model has been trained. It focuses on customization, delivery, and utilization of the pre-trained ML models. It enables customization of trained ML models for specific tasks, datasets, and domains, ensuring they meet diverse application needs. Additionally, it streamlines model delivery by managing ML model APIs and facilitating multi-model collaboration.
The purpose of this Recommendation defines functional requirements for adaptation of ML model, ensuring efficient customization, delivery, and utilization of trained ML models supported by cloud computing capabilities. It aims to streamline ML application development by reducing complexity, minimizing redundant training efforts, and enabling broader accessibility of ML-powered solutions.
|
|
Comment:
|
-
|
|
Reference(s):
|
|
|
Historic references:
|
|
Contact(s):
|
|
| ITU-T A.5 justification(s): |
|
|
|
|
First registration in the WP:
2025-03-19 11:28:30
|
|
Last update:
2025-04-23 18:32:30
|
|
