On 12 September 2019, the Artificial Intelligence Applied in Industry symposium at KU Leuven Campus De Nayer (Sint-Katelijne-Waver) addressed this question.
The symposiums centered on two main topics: analysis of time series data and applications of Deep Learning. More and more companies have started to collect more and more data; however, it is often not easy — especially for smaller companies — to derive useful conclusions from their data. Analysis of time series data (data collected over a specifically determined amount of time), is a key task for many companies, who wish to monitor their production or business processes in a (semi-)automated way. Deep Learning is the driving force behind recent innovations in computer vision, natural language processing, and many other domains. Both themes are directly relevant to today’s business needs.
Organized by the KU Leuven research groups Embedded and Artificially Intelligent Vision Engineering (EAVISE) and Declarative Languages and Artificial Intelligence (DTAI), the symposium presented results from two Technology Transfer (TETRA) projects: Intelligent Analysis of Time Series and Start to Deep Learn. Crucially, TETRA projects like these involve industrial-academic cooperation. Thus, a consortium of more than 20 industrial partners helped organize the symposium. In this case, a consortium of more than 20 industrial partners collaborated with KU Leuven. Both TETRA projects are funded by the Flemish Agency for Innovation and Entrepreneurship (VLAIO), ensuring that the most up-to-date technology is used to address the most industrially relevant use cases.
The event started with dr. Jan Van Haaren, head of Data & Analytics at SciSports, discussing the development of a new football performance metric. While previous metrics focused exclusively on technical and physical aspects, his team has come up with a new method that also addresses the mental pressure under which a player has to perform. They developed a machine learning model to estimate how much mental pressure the player possessing the ball experiences using a combination of match context features and the current game state. Given the extreme pressure put on top players, it can provide football clubs with a significant competitive advantage, using machine learning models to provide actionable insights for football clubs. He succinctly set the theme of the day: artificial intelligence and intelligent analysis benefits for real-life companies and situations.
Added value from temporal data
First on the agenda were results from the TETRA project Intelligent Analysis of Time Series, which ran from October 2017 to September 2019 and helped companies to derive added value from temporal data. The researchers investigated a number of case studies, aimed at a number of specific sectors, including machine manufacturing/maintenance, providers of complex ICT infrastructure and services. They attempted to use existing technology to analyze analogue, discrete and spatiotemporal time series. At the symposium, they presented results from three different case studies.
- oXya – Automated detection of anomalies in SAP
The first talk presented a collaboration between KU Leuven and oXya, an international company specializing in custom-made SAP solutions for other companies. As part of the project’s user group, oXya wants to automate the process of detecting anomalies in SAP server transactions for easier maintenance and faster response times. For this, KU Leuven researchers developed a framework that trains several models of the expected system behavior, and raises an alarm when the expected results differ significantly from the measured values.
- TenForce – leveraging past audits to predict future performance
TenForce helps customers in the utility sector audit maintenance sites. For them, the KU Leuven teams devised a system to optimize inspection rounds by leveraging past audits to predict future subcontractor performance. The system uses a Bayesian Network to model an audit, which is then used to generate a distribution over possible audit outcomes by means of Monte Carlo simulation. These distributions allow one to rank the open sites according to expected performance.
- Skyline Communications – Alarm sequence similarities
Skyline Communications specializes in network management and operational support for broadcast, telco, cable, satellite and mobile industries. One of their key challenges is to handle the large number of alarms that may occur, due to, e.g., poorly configured customer thresholds or cascading alerts. The KU Leuven researchers investigated a method for finding similarities between different sequences of alarms, based on graph isomorphism algorithms.
Deep learning technology for local companies
The second part of the symposium concerned the ongoing TETRA project Start To Deep Learn. The aim of this project, which runs from September 2018 to October 2020, is to facilitate local companies in the adoption of deep learning technology through hands-on workshops, seminars and real-life use cases on computer vision and AI. Again, participants showcased three case studies.
- Colruyt – occupancy degree and privacy
The first case study was conducted in collaboration with Colruyt. Like many modern companies, Colruyt allows employees to reserve meeting rooms and flex-desks. To ensure that these resources are used effectively, they want to monitor their occupancy, while still respecting their employees’ privacy. The KU Leuven researchers have proposed an approach to detect occupancy degrees on low-resolution omnidirectional video. To safeguard privacy, they are processing the data locally on an embedded platform and at reduced image resolution, achieving high accuracy for resolutions as low as 32px.
- Reliability computer vision systems
The second case study focused on the reliability of computer vision systems. Thanks to recent advances, computer vision is now routinely used in safety-critical applications. However, targeted attacks might be able to deceive such a system. Researchers are exploring this vulnerability by designing a specific patch that a person can hold in front of his or her torso to become undetectable by a state-of-the-art deep learning vision system. Using a method similar to training a neural network, they have already been able to automatically generate such an adversarial patch.
- Eurofins – Rapid, automated image comparison
The company Eurofins came to the research groups needing to develop a method for fast image comparison, for use in their newest digital TV testing solution. The solution involves testing the functionality of on-screen displays of new TVs image against the image from a camera pointed towards the TV. The KU Leuven teams compared several older image comparison techniques against Deep Learning, and found that the latter clearly outperformed the others, even without training.
Tangible benefits of AI research is a win-win
Ultimately, the symposium presented several successful instances of academic-industrial cooperation through funding schemes like these VLAIO-funded TETRA projects. The cases convincingly demonstrated the added value of such collaborations, both for companies (access to state-of-the-art science and technology) and researchers (access to real-world problems and datasets).
Moreover, the presentations illustrated diverse tangible benefits of using AI for companies. For example, Eurofins has already commercially implemented the findings of their case – that AI image comparison vastly outperforms surpasses previously used testing methods. The researchers demonstrated that companies can already use Deep Learning and intelligent analysis for efficiency monitoring, detecting SAP anomalies, and work planning based on previous performance. These are not just promises of technology to come, but evincible, realizable implementations from which companies could immediately profit.
Find out more and watch a short video of the symposium here.