Topological deep learning
Simplicial, cellular, sheaf, and Hodge-theoretic models for structured signals and higher-dimensional relations.
NeurIPS 2026 Workshop · Sydney only
From Geometry to Topology
Frontiers of Higher-Order Representation Learning.
Rethinking representation learning through higher-order structure in the foundation-model era.
40–50expected attendees in Sydney
10–15expected submissions
25–35planned PC members
Non-archivalworkshop contributions
A visual metaphor for moving from flat representations and pairwise relations toward multiscale, multimodal, and topological structure.
Workshop focus
Modern machine learning increasingly needs representations that preserve geometry, topology, group relations, and local-global structure across modalities and scales.
Central question
How should modern ML systems represent data when the relevant information is relational, multiscale, multimodal, constrained, or topological?
Why now?
The foundation-model era has sharpened this question. Token-, patch-, and pairwise-graph representations are powerful, but many scientific, visual, linguistic, and multimodal systems contain structure that is not naturally reducible to these forms.
Topics of interest
We welcome theoretical, methodological, empirical, benchmark, dataset, reproducibility, negative-result, and position papers.
Geometry and optimal transport
Geometric latent spaces, metric structure, Wasserstein methods, and topology-aware comparison of structured data.
Foundation-model pipelines
Structured latent spaces, multimodal reasoning, self-supervised objectives, and integration with modern encoders.
Scientific and biomedical AI
Physical systems, molecules, cells, clinical pathways, sensor interactions, and biological organization across scales.
Benchmarks and evaluation
Shared baselines, lifting procedures, datasets, evaluation protocols, reproducibility studies, and computational trade-offs.
Vision, NLP, and multimodal learning
Spatial layout, video interactions, language compositionality, code dependencies, grounding, and cross-modal correspondences.
Invited speakers
The speaker list is deliberately compact to preserve time for discussion, contributed work, posters, demos, and mentoring.
Vanessa Robins
ANU, Australia
Computational topology, TDA, persistent homology.
Bastian Rieck
University of Fribourg, Germany
Topological ML and robust multi-scale analysis.
Katharine Turner
ANU, Australia
TDA, metric geometry, and statistical topology.
Marco Cuturi
Apple, CREST - ENSAE
Optimal transport and Wasserstein geometry.
Draft programme
A one-day workshop with invited talks, contributed spotlights, posters, demos, a panel, tutorial, and mentoring tables.
09:00–09:15
Opening
09:15–10:00
Invited Talk 1
10:00–10:30
Contributed spotlights · 3–4 talks
10:30–11:00
Coffee and poster/demo session 1
11:00–12:00
Invited Talk 2
12:00–13:00
Panel · When do higher-order models genuinely outperform pairwise and Euclidean alternatives?
13:00–14:00
Lunch and mentoring tables · Students and early-career researchers
14:00–14:45
Invited Talk 3
14:45–16:45
Tutorial
16:45–17:00
Closing remarks and next steps
Call for papers
We invite non-archival submissions, including extended abstracts, short papers, position papers, benchmark and dataset reports, reproducibility studies, and negative results. Student-led and interdisciplinary submissions are especially encouraged.
Submission deadline
29 August 2026
Decisions finalised
26 September 2026
Author notification
29 September 2026
Submissions will be managed through OpenReview. Accepted papers will be made available as non-archival workshop contributions, following NeurIPS workshop policy.
Organising team
The team spans topological deep learning, geometric representation learning, optimal transport, multimodal sensing, NLP, computer vision, and applied AI.
Mathilde Papillon
UCSB · Topological deep learning.
Colleen Farrelly
Post Urban Ventures · Mathematical and applied AI perspectives.
Ali Zia
La Trobe University · Machine learning, computer vision, and topological deep learning.
Abdelwahed Khamis
CSIRO · Optimal transport, topological deep learning, and multimodal sensing.
Usman Naseem
NLP and representation learning.
Mustafa Hajij
University of South Florida · GRaM Workshop and TDL Challenge organiser.
Advisory contributors
Wei Xiang, La Trobe University. Zeeshan Hayder, Google Australia.
Programme committee plan
We plan to recruit 25–35 PC members, with approximately three reviews per submission and a maximum of three papers per reviewer.