Hong-Linh Truong

Context: The rapid adoption of AI in everyday life and critical domains increases the need for explainable AI systems. However, software engineering research provides limited understanding of how explainability is perceived and implemented in practice, and how organizations can comply with the transparency requirements of the EU AI Act (AI Act). Objectives: This study has two objectives: first, to understand how industry experts perceive the explainability of AI systems, and second, to demonstrate how explainability can address the transparency requirements of the AI Act. Method: We conducted a case study in a consulting company by interviewing industry experts and analyzing their documents on the AI Act, trustworthiness, and sustainability of AI systems. Results: The case study revealed that one of the key reasons for explainability is to increase users’ understanding not only of the operations of the AI systems but also of its risks and limitations. The industry experts also emphasized that explainability is important to ensure accountability and human agency. One of the key challenges is that customers are not interested in investing in the explainability of AI systems. During AI system development, the lack of continuous end-user testing is also a significant explainability challenge in practice. To support organizations in addressing the transparency requirements of the AI Act, the paper proposes an approach that consists of a set of principles, practices, and guiding questions. Conclusion: For researchers, this study offers insights into the reasons and challenges of defining and implementing the explainability of AI systems in practice. For practitioners, the paper describes a lightweight approach, as recommended by industry experts, for implementing the transparency requirements of the AI Act as explanations of AI systems.

Multi-modality machine learning (ML) systems are increasingly deployed in edge environments, such as smart building and autonomous robotic applications. These systems enable analysis of complex subjects by performing ML inferences on multiple data sources provided by Internet of Things (IoT)-enabled devices. However, operating such ML systems for multi-tenant applications presents a runtime orchestration challenge to optimize the execution of multiple inference tasks across modalities. Analysis requests from different tenants often rely on different data sources with distinct quality requirements, resulting in a wide range of conflicting optimization objectives for runtime inference orchestration. The diversity of ML models and their performance variations further complicate the orchestration, particularly in scheduling inference tasks and distributing inference workloads on a heterogeneous edge system. This article addresses these challenges by introducing an adaptive orchestration that manages inference task scheduling for multi-tenant applications in a multi-modality ML system. Our orchestration supports a complex tradeoff between quality refinements and inference capability of inference services while scheduling multiple inference tasks under dynamic time constraints. The orchestration employs an efficient mechanism for selecting instances of inference services to distribute inference workloads across a heterogeneous edge cluster. This mechanism allows the orchestration to leverage cross-modal information to refine inference quality following the tenant objectives. Furthermore, our orchestration considers the performance variation caused by resource contentions, runtime failures, and explainability overheads while distributing workloads. Extensive experiments on two real-world edge applications in a common heterogeneous edge testbed demonstrate that our orchestration can improve inference accuracy by up to 10% and reduce the late response ratio by up to 18% compared to existing approaches within our multi-tenant edge scenarios.

The Internet of Things (IoT) is becoming increasingly ubiquitous, acting as an important source of real-time data for various applications. By allowing data exchange between various parties along the IoT devices-Edge-Cloud computing continuum, the larger societal benefits of the IoT can be achieved. Assuring security and fostering confidence for IoT data sharing, however, is one of the biggest obstacles. Sharing real-time data originating from connected devices is crucial to real-world intelligent IoT applications, i.e., based on artificial intelligence/machine learning. Such IoT data sharing involves multiple parties for different purposes and is usually based on data contracts that might depend on the dynamic change of IoT data variety and velocity. We aim to support multiple parties (aka tenants) with dynamic contracts based on the data value for their specific contextual purposes. This work addresses these challenges by introducing a novel dynamic context-based policy enforcement framework to support IoT data sharing (on-Edge) based on dynamic contracts. Our enforcement framework allows IoT Data Hub owners to define extensible rules and metrics to govern the tenants accessing the shared data on the Edge based on policies defined with static and dynamic contexts. We have created an edge-centered architecture that enables multi-tenant use cases with tenant-specific application deployment and IoT-context-based data sharing on edge servers. Our proof-of-concept prototype for sharing sensitive data such as surveillance camera videos has illustrated our proposed framework. The experimental results demonstrated that our framework could soundly and timely enforce context-based policies at runtime with moderate overhead. Moreover, the context and policy changes are correctly reflected in the system in nearly real-time. We have addressed the need to enable multi-parties IoT (data) resources to be shared based on contracts, especially with dynamic IoT contexts, for tenant applications on the edge to allow their closer access to data.

Given a large-scale mobile network with a variety of equipment and radio access network technologies for an approximate 20 million subscribers, there are many types of data that can be used for big data analytics and machine learning (ML) tasks for network operations, monitoring, and optimization. However, a variety of data is measured, collected, and propagated through numerous complex data and software systems. Thus, people, software components, and data-driven operations for big data and ML pipelines face great challenges in dealing with data quality impacts. Data quality related problems occur and are propagated through complex operations involving different types of data, people, software components, and analytics that cannot be solved purely through data quality engineering. This article discusses our TENSAI framework, as a practical and responsible observability for ensuring data quality in such a mobile network. TENSAI focuses on methods of communication, strategy specifications, and data quality engineering for diverse types of data and analytics among different types of operations. TENSAI presents techniques for capturing and communicating causes/effects of data quality problems clearly to all relevant stakeholders, developing data quality-aware adaptation strategies for actions on data that can be integrated into analytics processes, and engineering data quality awareness in software and data pipelines. Thus, TENSAI supports full visibility of data quality problems and impacts among related systems to empower the utilization and adaptation of data analytics for different types of operations. We will illustrate our TENSAI with several real-world data types, pipelines, and cases based on a real-world mobile network.

The increasing development and utilization of Large Language Model (LLM) services have demonstrated many benefits in different contexts. However, LLM services are mainly available in the public cloud and require huge computing resources to operate, thus not accessible to many companies, organizations or communities with constrained resources. While research efforts have concentrated on LLMs quantization for resource-constrained computing environments like edge devices, to democratize the availability of LLM services as utilities for such communities requires much more than the optimization of LLM models. In this paper, we introduce CuLao – a framework for constructing utilities from LLMs in resource-constrained environments. Our framework focuses on key requirements of resource-constrained companies, organizations and communities by enabling the provisioning and coordination of LLMs as utilities, based on the availability of open-source LLMs.

The growing complexity of end-to-end Machine Learning (ML) serving across the edge-cloud continuum has raised the necessity for runtime explainability to support service optimizations, transparency, and trustworthiness. That involves many challenges in managing ML service quality and engineering runtime explainability based on ML service contracts. Currently, consumers use ML services almost as a black box with insufficient explainability for not only inference decisions but also other contractual aspects, such as data/service quality and costs. The generic explainability for ML models is inadequate to explain the runtime ML usage for individual consumers. Moreover, ML-specific metrics have not been addressed in existing service contracts. In this work, we introduce a novel contract-based runtime explainability framework for end-to-end ensemble ML serving. The framework provides a comprehensive engineering toolset, including explainability constraints in ML contracts, report schemas, and interactions between ML consumers and the components of the ML serving for evaluating service quality with contract-based explanations. We develop new monitoring probes to measure ML-specific metrics on data quality, inference confidence, inference accuracy, and capture runtime ML usage. Finally, we present essential quality analyses via an observation agent. That interprets ML inferences and evaluates contributions of ML inference microservices, assisting ML serving optimization. The agent also integrates ML algorithms for detecting relations among metrics, supporting constraint developments. We demonstrate our work with two real-world applications for malware and object detection.