Of course, what happens at a TechDays party stays at a TechDays party 😉 But we can share some of the AI highlights with you. So here we go, get ready for a glimpse of our orange-tinted AI fiesta! Check out the aftermovie:
We were joined by the following speakers who shared real AI case studies from their companies:
As AI advances, the projects to implement it are getting bigger. Realer. And that growth can really escalate the costs in some cases. Shahin Shahkarami (IKEA) presented an exciting talk about how they’re using Q Learning to revolutionise shopping experiences on ikea.com. “Q learning… isn’t that expensive?” asked an audience member during the Q&As.
It wasn’t the only time the topic of cost came up. Many of the companies presenting were talking about really resource-heavy projects, and how these were going to make a big impact in their business. We’re not talking AI hype here; these are big investments with solid reasoning behind them. Some of the projects already are making a big impact. For instance, as Carlo Bruno (Adyen) explained, Adyen works with really enormous numbers of transactions. Machine learning helps them process transactions more quickly and smoothly, which is essential for their business.
There were a number of projects angling for the ‘biggest’ bragging rights at this edition of TechDays 😅 an, and not just when it comes to investments. And size really does matter when you’re training AI models.
We’re seeing some companies working with really huge amounts of data now, especially as the processes of data collection and cleaning are becoming more streamlined. And that’s awesome, because more quality data means more accurate models and better results. Other companies are grappling with enormously high cardinality, or models that require 8 hours to complete one iteration. It’s big stuff.
Speaking of big projects, Ronald den Elzen and Surajeet Ghosh (The HEINEKEN Company)shared a sneak peek of Heineken’s plans for integrating AI. It’s super ambitious, and it’s still a work in progress, but the way they envision connecting their divisions and processes all over the world is truly epic. And the strategy behind it is a lot more than just AI hype.
For some companies and organisations, scaling up means considering more than just the technical factors. Although the EU AI Act is still months away from becoming reality, privacy and legal issues exist already for those working with images. We heard from Hezha MohammedKhan (Zero Hunger Labs), who is developing a method for detecting malnutrition in children using images. Naturally, you can’t just gather thousands of images of children without properly checking that you’re following the rules for permission, storage, and use of those images. For Zero Hunger Labs this has slowed down the development of their model, but they’re still on track to succeed.
Meanwhile, we heard from Robert Musters (Nederlandse Spoorwegen) that NS has already developed a project to automatically protect privacy in images captured at NS stations. By using AI to detect images where people are present and blur them out, they can gather the useful visual information they need (images of trains and station equipment, for example) much more quickly and efficiently, and still be compliant with privacy requirements.
Daan Odijk (RTL) explained that their media company faces quite a different issue when using AI with images. Not only have they been grappling with the question of whether they’re allowed to use a particular AI-generated image, they also need to be careful to protect their own intellectual property when uploading images into an AI tool. One important lesson they shared is the importance of checking the terms and conditions of each generative AI tool, because they’re not all the same. In spite of the legal uncertainties, they’ve been doing a lot with AI. Like creating these really cool promos with Rivella!
For many of the companies who presented at TechDays, their AI projects stay mostly in the digital realm. Sure, the data comes from the physical world, and their models can have an impact on real world events. But it’s still mainly a digital endeavour. So it was really interesting to see some cases where the AI model needs to be tested or reviewed outside of a computer.
For instance, Deirdre Douma and Noortje van Genugten (Albert Heijn) joined us to talk about AH's Dynamic Markdowns. This innovative approach to discounting products close to expiry could seriously reduce food waste, but getting it right is complicated. Every change could impact user behaviour in-store — and not just right now, it could have an impact tomorrow, or next week. As a result, they’re actually testing it on the ground, in the physical world, to find the optimal balance of discounts to reduce food waste.
ING’s project to automate lending decisions faces a different challenge, as Anil Panda and Artur Usov (ING) explained. Their model tries to predict who will default on a loan, and the only way to validate that in the real world is to wait to see who actually does default on a loan. Testing it means months of giving out loans not knowing if the model was approving things it shouldn’t be. To mitigate this risk, ING has a whole set of checks and reviews from independent parties to ensure that the model performs as expected (and within regulations).
Thanks to all our speakers and attendees for joining! We'll be back soon with our next TechDays event, so stay tuned for more awesomeness 😎
