Laboratory Life
Prof. Hong-Linh Truong, from Aalto University in Finland, gave a talk. Date and Time: 15:00 - 16:00, 15th December, 2022 Place: Social Intelligence Labratory, BKC, Ritsumeikan
Summary
Recently, the number of machine-learning-based Web services (MLaaS) has been increasing. Unlike the existing services, the cores of these services are blackboxes and the behavior varies according to the versions of training data. In combining multiple MLaaS, even if those interfaces are compatible, we may fail to invoke them because the difference of the training data can change output of each service. This talk introduced QoA4ML, a specification language for MLaaS proposed by Prof. Truong, and monitoring tools. The presentation was translated into Japanese by a voice translation service combining a speech recognition service with machine translation service, typical MLaaS.
New Frontiers on Supporting Contracts for Machine Learning Services: The QoA4ML Approach
Abstract
Many complex software now utilize machine learning (ML) inference features offered by ML services as a part of the software. Such ML services encapsulate advanced ML models and offer service APIs and they are the key element of the ML as a Service (MLaaS) paradigm. Unlike other types of the services in the edge and cloud, the core of ML services is based on ML models, which have powerful abilities but exhibit several contractual problems for service providers and consumers in terms of operation deployment, provisioning and regulation governance due to the ML inner blackbox and data effect characteristics. This calls for new frontiers in service, software system and ML communities to address concerns about contracts in ML services. Unfortunately the problem is currently tackled in a silo manner (e.g., software system focused on ML service and pipelines provisioning and management whereas ML community is on ML model performance. In our view ML services introduce several complex quality-related concerns that need to be tackled together. In this talk, we will present our approach in the QoA4ML framework. We support the identification and specification of concerns covering different aspects of data, ML models and services. Based on that we develop policies for ML service contracts. We will discuss our current prototype, which focuses on MLaaS for three party interaction models and ML services in edge-cloud continuum.
Bio
Hong-Linh Truong is a tenured associate professor at the Department of Computer Science at School of Science, Aalto University, Finland. He leads the AaltoSEA Group on Systems and Services Engineering Analytics. He is also a Priv.-Doz. (Adjunct Associate Professor) at the Faculty of Informatics, TU Wien. He received a PhD (2005) and a Habilitation (2013) from TU Wien, Austria. Among his visiting positions, he was a visiting scholar at University of California, Irvine and at University of Southern California and had short visits to National Institute of Informatics (NII), Japan and lectured in Fudan University and HoChiMinh City University of Technology as a guest professor. His main research interest focuses Systems, Software, Data and Service Engineering Analytics by developing novel techniques and tools for monitoring, analyzing, and optimizing functions, performance, data quality, elasticity, and uncertainties associated with systems, software, data and services. His research has been applied to: Monitoring, Analysis and Optimization Techniques for Programs, Data and Systems; Parallel, Grid and Cloud Computing, and IoT; Big Data; Data Service Models and Analytics; Socio-technical Services Engineering; and Elastic Computing. Furthermore, he is interested in (free) ICT solutions for (under) developing countries. He had delivered several invited talks and he published more than 200 refereed papers in books, conferences/workshops and journals (Google Scholar citations=6544, h-index=41). He (co)receives an outstanding paper award, seven best paper awards, one best paper run-up award, and one best poster award. (https://users.aalto.fi/~truongh4/).