2026-09-10 –, Room 1 (170)
In a time where important tasks and decisions are handled by large and opaque AI models, it’s easy to lose touch with the algorithms behind them. Do we still recognise the building blocks? Can we still reason about how models work?
This talk invites you into a time machine to rediscover algorithms you may have missed or forgotten about. By seeing the beauty and imperfections of past technological solutions, we hope to gain a clearer view on evaluating models in the current age.
Drawing from over a decade of academic and industry experience, four 20th-century classical Machine Learning families are presented based on real-life use cases. With a minimum of formulas and technical details, we’ll see what made the ideas behind them so compelling, what we can learn from them, and if they are still relevant today.
The format of the presentation is largely conceptual and with an emphasis on distilling applicable insights.
Who needs to attend this? It’s primarily for data scientists (practitioners and researchers at all levels), but really also for everyone interested in the history and philosophy of AI. Basic knowledge about machine learning is beneficial.
Best available Python implementations of the discussed methods are provided. Code with working examples will be shared in a Github repo.
You will leave this talk with a richer appreciation of classical ML, with valuable modelling insights, and you’ll have what’s needed to try your hand at a few funky classics yourself.
In an age dominated by generative AI, it’s easy to lose touch with the foundational building blocks of modelling and algorithmic problem solving.
This informal talk is not an introduction to classical Machine Learning (ML), nor a mathematical deep-dive into the details of learning algorithms. Instead, it’s a fun-yet-serious history of science tour honouring the life cycles of lesser known classical ML methods.
This should be interesting to new generations of data scientists, who may not have had the chance to learn about these algorithms yet, and older generations as well. Both are invited to reflect on whether today’s models are in every way better than the oldies.
We’ll walk through the high-level working of a few selected algorithms. Each will be illustrated with a real-life use case. In addition, valuable, universal modelling insights are highlighted. This then leads to surprising connections and verdicts about these methods.
Python code snippets and simulation results are shown throughout the presentation, and a link to a public Github repo with notebooks and literature references will be shared. The talk itself is not a full demo or tutorial, however.
The takeaways of this talk flow naturally from the benefits of studying classical ML. It’s not just about understanding how algorithms learn from data, but also about expanding your toolkit and ability to see cross-links. And finally, we might even recognise cases where the original is the better choice.
Structure:
Introduction — 5 min
Algorithm 1: Kernel Machines — 5 min
Algorithm 2: Genetic Programming — 5 min
Algorithm 3: Orthogonal Matching Pursuit — 5 min
Algorithm 4: Hierarchical Mixture of Experts — 5 min
Final insight and Q&A — 5 min
As an experienced data professional, Christiaan’s interests range from data science and predictive modelling, to the engineering behind AI and ML systems.
After specialising in hydrodynamics, he worked for several years at Deltares as a consultant on the design of hydraulic structures in The Netherlands. His PhD research on vibrations induced by turbulent flows led him to adopt machine learning techniques early on.
Subsequent data science employment in the UK included Internet-of-Things applications at Centrica, and payments optimisation — most recently at Worldpay. Having founded and managed both data science and MLOps teams, he is now freelancing in Amsterdam while also working on personal projects.
Christiaan holds an MSc in Civil Engineering from the TU Delft and a PhD in Computational Science from the University of Amsterdam.