A paper from the Human-Centred Artificial Intelligence Lab has been accepted to the main track of Neural Information Processing Systems (NeurIPS) 2026, one of the main conferences in artificial intelligence. The team will present the work in Sydney in December 2026.
The paper is Hypernetworks for Dynamic Feature Selection. Dr Javier Fumanal Idocin, a YUFE MSCA Postdoctoral Fellow, is first author. Prof. Javier Andreu-Perez is a co-author, with Raquel Fernandez-Peralta. The preprint is on arXiv as 2605.12278, and the code is at github.com/Fuminides/hyper_DFS.
Dynamic feature selection acquires features one at a time, under a budget. Existing methods do not always find the best solution as the number of possible acquisition paths grows. Hyper-DFS uses a hypernetwork to generate a classifier for a particular feature subset when it is needed. A Set Transformer encoding keeps similar tasks close together in that conditioning space.
On the benchmarks reported in the paper, Hyper-DFS was best or second best against the compared methods on synthetic and real tabular data, and best or second best across the image datasets tested. It also generalised better, without further training, to feature subsets it had not seen. The University notes possible applications where collecting each extra piece of information costs time and money, including medical diagnosis.
The University of Essex published the announcement on 2 October 2026.