6 recordings, newest first by talk or conference date. Conversations use the episode release date.
TaDA Workshop · VLDB 2026
Research talk · 11:19
Hyperdimensional computing for structured querying on tabular data
Paper authors: Sebastián Bugedo, Stijn Vansummeren
Abstract · summary
Vector embeddings can retrieve similar table rows without revealing whether the nearest result actually satisfies a structured query. This work uses hyperdimensional computing, specifically holographic reduced representations, to encode rows and evaluate selection and projection in vector space. It derives expected similarity scores that support principled retrieval thresholds, including detection of queries with no matching rows. Experiments compare the approach with EmbDI on two real datasets and examine equality, non-equality and attribute projection.
Can We Outsource Thinking? AI, Education, and the Future of Knowledge Work
In conversation: Frank Neven, Kris Peeters
Episode summary
Frank Neven and Kris Peeters discuss what AI can automate in data engineering, software development, research and education, and why understanding remains important when reasoning is delegated to AI. The conversation covers agentic systems, database research, structured data and semantics, human–AI collaboration and personal knowledge management. It considers how people can work effectively with AI while continuing to learn and judge its outputs.
A Logical View of GNN-Style Computation and the Role of Activation Functions
Paper authors: Pablo Barceló, Floris Geerts, Matthias Lanzinger, Klara Pakhomenko, Jan Van den Bussche
Abstract · summary
MPLang provides a declarative view of message-passing computation in graph neural networks. The study characterises the fragment without activation functions through walk-summed features and establishes conditions under which eventually constant activations have equal expressive power. With linear layers available, it proves that unbounded ReLU can express numerical queries that bounded, eventually constant activations cannot. These results clarify how the choice of activation function affects the computations a GNN can express.
Models are data too: Towards expressive querying of machine-learning models
Speaker: Jan Van den Bussche
Abstract · summary
Machine-learning workflows produce many models that need to be stored, managed and queried as data. This keynote explores how expressive, declarative query languages could support that task. It surveys existing approaches and open research questions, connecting model querying with the broader challenge of querying program code. The aim is to encourage database research on models as a distinct kind of data.
Robustness Against Read Committed: A Free Transactional Lunch
Speaker: Frank Neven
Abstract · summary
Read committed often increases transaction throughput, but can permit behaviours that serializable execution rules out. This keynote examines when an application's transaction programs are nevertheless guaranteed to behave serializably at the weaker isolation level. The associated survey reviews sound and complete tests for robustness against multiversion read committed, explains how the analysis extends from individual transactions to templates and programs, and outlines open research questions. The aim is to identify workloads that retain serializability without paying for stronger isolation.
Robustness against Read Committed for Transaction Templates
Paper authors: Brecht Vandevoort, Bas Ketsman, Christoph Koch, Frank Neven
Abstract · summary
Transaction templates model database programs so that their behaviour can be analysed before execution. This work develops tractable tests for deciding whether every schedule allowed by multiversion read committed is serializable. The approach recognises larger robust subsets of workloads than earlier methods and explores how promoting selected reads to updates can make additional workloads robust. Experiments show how these guarantees can improve throughput by allowing workloads to run safely at a weaker isolation level without changing database internals.