'Artificial intelligence involves planning, deciding and reasoning'. Takeaways from the Beltug N-sight, 19 November 2019


Companies are finding it hard to grasp the tangible potential of Artificial Intelligence and Machine Learning, and to find ways to cash in on the opportunities.  The technology remains rather enigmatic.


We organised this session to clear a bit of the haze, and to discover how AI and ML initiatives are playing out in business, public entities and academia. We heard two real-life cases: on rolling out data mining and AI at FPS Finance, and on how SD Worx uses AI and big data to integrate new services. We discovered some specific ways to use machine learning to enhance digital experiences, from Dropsolid. And we found out what the KU Leuven’s Declarative Languages and Artificial Intelligence (DTAI) research group has been working on in the area of industry.


The presentations from the N-sight are available for our members (after login).



Case: Towards an analytics-driven organisation at FPS Finance Belgium


Dierk Op 't Eynde, Senior Advisor Strategic Support at FPS Finance and Board Member of the Royal Statistical Society of Belgium, opened the session enthusiastically. FPS Finance has been restructured thoroughly in the last few years, including a 40% reduction in personnel and a 70% decrease in regional offices. 


To optimise processes, guarantee compliance and keep focused on the high risks, FPS Finance took on AI and data mining in 2010. Data mining teams in several departments and administrations have projects running (slide 5).


Dierk illustrated one use case about the Special Tax Inspectorate, describing how data mining has increased accuracy, with fewer physical controls.  VAT carousel fraud has been reduced by 90% in Belgium, also thanks to data mining and AI.


We also learned about a use case on tax-on-web.  The complexity of the personal income tax in Belgium has only increased, and this trend is likely to continue.  The original tax-on-web application was not developed for this level of complexity.  To tackle this challenge, FPS Finance created a digital twin which, using advanced analytics, enables government negotiators to simulate the impact of new regulations, for instance. Dierk demonstrated an actual simulation of possible scenarios for us.



How to use practical Machine Learning in improving the digital experience


Dominique De Cooman, CEO of Dropsolid, then shared his insights on how Machine Learning can improve the digital experience of your clients or your colleagues. We were introduced to two use cases. The first was 'search relevancy’. When Amazon applied this to visitor searches, its conversion rate shot up 6x: from 2% to 12%.  'Personalised messaging' is the second use case (slides 19 and 20).


Dominique then share a real-life experience: its customer, hospital group GZA, increased search result accuracy by 91% after only two weeks.  While it isn’t really innovative, it is an easy opportunity: conversion rates for visitors to your website go up when the search results are accurate, and hopefully, that translates to more purchases or a longer visit.


So what do you need for this? An innovative mindset, for starters, said Dominique.  Consider a separate team that focuses on innovation.  Cloud capabilities is another requirement - this gives you scalability.



Case: How AI helps us give value to our data


Peter Van Ostaeyen, Manager Innovation & New Technologies at SD Worx started his talk with a quote from Elon Musk: "If you're not failing, you're not innovating".  However, company Boards generally want to know the ROI of an idea.  This is particularly hard to calculate for AI.  Luckily, Peter explained, SD Worx has an innovative spirit.


In a first stage, a think tank defined 3 areas where AI will bring added value:


  • Anomaly detection
  • Data field learning
  • NLP deployment


The goal of the data field learning, is to predict outcomes and to look for a deeper understanding. Through data field learning, SD Worx learned that in retail, absences peak in the end-of-year period (slides 7-9).


Peter showed us how running an algorithm on the absence data gives actionable, predictive models. What did SD Worx conclude from this data field learning project?


  • End of year absence levels are fuelled by fixed staff being on holiday (or sick), an increase in temp workers, and instore traffic peaks.
  • Retail companies with flu vaccination programs save money.
  • Hiring store personnel beyond a 10km perimeter increases absences in peak moments and amongst certain age groups.


Peter shared more use cases, as well (see slides).



ML and AI: From research to application 


Our final exposé of the day came from Wannes Meert, Research Manager, Artificial Intelligence research group at KU Leuven.  There exists no real definition of Artificial Intelligence, he started. But he did gather a few descriptions for us (slide 3).


AI requires 2 important features, he continued: 'self-inspection' and 'self-adaptation'.  Adding these to machine learning/data mining, together with 'reasoning', 'knowledge representation' and 'online', gives you 'Artificial Intelligence'. Then he outlined what we can and cannot do with AI today (slide 5).


Wannes took us through machine learning, explaining that a machine learns if it is able to:


  • improve its performance
  • on a specific task
  • with experience


Machine learning is extremely good at interpolation, yet very bad at extrapolation, he described.  But even with limited resources, big tasks can indeed be automated (slide 17).  Keep in mind the limitations though:


  • If training data/labels are hard to collect, you need HW/SW infrastructure
  • If large quantities of data are required, you need HW/SW infrastructure
  • If it is not a fixed model, you need a complicated pipeline.


And datasets — not algorithms — might be the key limiting factor to the development of human-level artificial intelligence. If you have no labels, change the task, Wannes advised, and use anomaly detection to detect deviations from the dataset.


His takeaways for machine learning:


  • Most successes have been achieved in supervised learning
  • Supervised learning requires labels
  • There is no single good model
  • Learning is interpolation, not extrapolation
  • Think about the actionability of the algorithm’s outcome
  • Think about closing the loop and provide feedback


Then he turned to Artificial Intelligence, which is all about reasoning, rather than just patterns (slides 27-34).


His conclusions:


  • Artificial intelligence involves planning, deciding, reasoning
  • A combination of techniques and knowledge sources is required
  • The building blocks are general but the pipeline is application-tailored
  • Interdisciplinary collaboration between AI-engineers and domain experts is crucial















Dear visitor,

Access to more information about this topic and/or to download the paper is easy and fast, but exclusively for Beltug members (just login to get access).

Beltug gathers a lot of information. Here you find the advantages of Beltug membership

The Beltug Team

Click here to login

>>> Back to overview