ELLIS Machine Learning Insights Seminar Series
The ELLIS Unit Delft hosts the Machine Learning Insights Seminar Series. Presentations showcase research underway by members of the ELLIS Unit Delft, as well as invited guest speakers. Seminars are open to anyone interested in state-of-the-art developments in the fundamentals of machine learning, including researchers of all levels from TU Delft and other institutions. The series is centred on exchange and interaction; we encourage participation and discussion from all attendees!
When & where: seminars take place monthly throughout the academic year. All seminars are hybrid. See below or check the event feed on the homepage for upcoming seminars.
-
Tuesday 1 April 2025, 16:00 @ Echo Arena
Abstract
In many domains of science and society we are interested in making good decisions. For instance, should we use drug A or drug B to treat a disease, will this advertisement lead to higher spending, or will this new education intervention lead to better learning outcomes? Statistical machine learning does not directly address these questions, but perhaps we can use it as part of a solution to do better, more trustworthy causal inference to answer such questions. In this talk, I will cover various building blocks my research group investigates to address this, starting with a discussion of the inadequacies of prediction, to the difficulties of causal assumption checking and the challenges of personalized decision support using experimental data. We end with some current open problems and future directions of investigation that our “Safe Causal Inference” consortium will work on in the coming years.
-
Tuesday 4 March 2025, 16:00 @ Echo Arena
Abstract
As the Artificial Intelligence of Things (AIoT) transforms our digital landscape, the fusion of edge computing and communication networks creates an urgent need for more efficient data processing solutions. Modern systems face the dual imperatives of minimizing latency and power consumption—critical not just for user experience but for environmental sustainability. Our research leverages neuromorphic computing's brain-inspired principles to address these challenges to have delivered three generations of dynamically sparse delta neural network accelerators, DeltaRNN, EdgeDRNN, and Spartus, achieving over 40× speedup in speech recognition while maintaining power efficiency and accuracy. Building on these neuromorphic principles, we extend our research to eye tracking for extended reality and radio frequency (RF) signal processing, specifically targeting non-linearity correction in wideband RF power amplifiers crucial for emerging 6G and WiFi 7 technologies. Through OpenDPD, the first open-source digital pre-distortion (DPD) framework, we enable systematic training and benchmarking of AI-based DPD algorithms. Additionally, our mixed-precision AI-DPD approach (MP-DPD) reduces power consumption by 3× in processing near-GSps-level data rate RF signals. Looking ahead, we envision these advances helping to create ultra-efficient, intelligent edge devices that seamlessly integrate with next-generation communication networks.
-
Tuesday 4 February 2025, 16:00 @ Hilbert Room (Informatica)
Abstract
Arguably, a desirable feature of a learner is that its performance gets better with an increasing amount of training data, at least in expectation. Surprisingly, such behavior is not guaranteed for various standard learners. This talk makes the basic setting precise and gives some illustrations of this so-called nonmonotonic (or not-so-smart) behavior. In particular, it dwells on some recent results obtained, both theoretical and empirical, for a technique ubiquitous in data science: k-means.
More information and link to join online: https://www.tudelft.nl/ellis-1/ellis-machine-learning-insights-seminar-marco-loog-tu-delft
TU Delft PhD Candidates
ELLIS ML Insights Seminars are eligible for Discipline Related Skills graduate school credit with the ‘Form for earning GSC for TU Delft AI(-related) seminars’. Per 4 seminars attended, PhD candidates can earn 0.5 GSC up to a total of 1.5 GSC. The PhD candidate is responsible for collecting the confirming signature of participation from the organiser or speaker per seminar on this form. Check with your local Faculty Graduate School (FGS) if they offer this option, and if your supervisors accept our seminars as part of your Doctoral Education (DE). If you already have a form (e.g., from AI Lunch series), don’t forget to bring it with you.