Active Learning: A powerful tool for Accelerated Acquisition of Knowledge
Industry 4.0 is the automatisation of traditional manufacturing and related industries, using modern technologies and controlling the industrial value chain. The increasing digitalisation of manufacturing has accelerated the flow of information. Technologies such as Cyber-Physical Systems (CPS), Industrial Internet of Things (IIot), and Artificial Intelligence (AI) add value to Industry 4.0 value chains. In this way, products and means of production are networked and can "communicate," enabling new ways of production, value creation, and optimisation in real-time.
One of the project goals of STAR is to research and integrate leading-edge AI technologies such as active learning systems, explainable AI, human-centric digital twins, and much more, to allow the safe deployment of sophisticated AI systems in production lines. The Jožef Stefan Institute main competencies in the STAR project are in the areas of data analytics and machine learning. We develop methodologies and approaches for active learning. Active learning (AL) is usually the natural approach to provide human-in-the-loop (model requiring human interaction) functionalities in advanced AI systems. Typically, AL attempts to improve learners' performance by asking questions to an expert to obtain labels for data instances. Since users are often reluctant to provide information and feedback, AL is used to identify a set of data instances on which the provided users' input conveys the most valuable information to the system. In the decision-making process in manufacturing, AL can also be implemented in recommender systems. In such cases, it tackles obtaining high-quality data that better represents the user's preferences and improves the recommendation quality. The ultimate goal is to acquire additional feedback that enables the system to generate better recommendations. Collecting feedback from forecast explanations can be realised with a framework of three components: a forecasting engine, an explanation engine, and a feedback loop to learn from the users. We extend this approach to collect feedback from forecasts, forecast explanations, and decision-making options we recommend to the users.
We have developed a system that can acquire and encapsulate complex knowledge. The system is based on semantic technologies, considering ontology concepts that are generic and ported to multiple use cases. It integrates demand forecasting models, explainable AI (XAI), a decision-making recommender system, and a knowledge graph. The components mentioned above are used to develop decision-making workflows displayed through an interactive user interface. Feedback is collected from users regarding forecasts, forecast explanations, and decision-making options shown to the users.
The system requires at least eight components:
- Database - Stores data from manufacturing plants.
- Knowledge Graph - Stores data ingested from a database or external sources and connects it, providing semantic meaning.
- Active Learning module - Aims to select data instances whose labels are expected to be most informative to the system and thus help enhance the AI model's performance.
- AI model - Aims to solve a specific task relevant to the use case, such case.
- XAI Library - Provides some insight into the AI model's rationale to produce the output for the input instance considered at the task hand.
- Decision-Making Recommender System - recommends decision-making options to the users.
- Feedback module - collects feedback from the user and persists it into the knowledge graph.
- User interface - provides relevant information to the user through a relevant information medium.
The current work presents a system's conceptual design to acquire and encapsulate complex knowledge using semantic technologies and AI. The system is demonstrated on a demand forecasting use case in manufacturing. The methodology can be extended to several use cases in manufacturing. The system provides forecasts, forecast explanations, decision-making options, and the capability to provide implicit and explicit feedback. It enables the development of an active learning module that can improve data collection by identifying promising data instances that, when labeled, are expected to be most informative to the system. Our future work in the STAR project will focus on implementing an active learning module and explore recommender systems that learn from data to provide decision-making options to the users. Stay with us and follow the latest news!
By: Patrik Zajec, Jože M. Rožanec, Inna Novalija, Blaž Fortuna, Klemen Kenda, Dunja Mladenić and Kim Sevšek / JSI