I'm honored to become a UKRI Future Leaders Fellow! I will push the frontiers in topological deep learning and 3D vision for the next 4+ years. Several PhD and Postdoc positions available. #UKRIFutureLeaders @UKRI_News #FLF @imperialcollege @ICComputing.🔗

very special for us, too. 🙏.

RT @abursuc: The @CVPR reception w/ a concert of a band of researchers to and audience of thousands of researchers is quite a special event….

There is geometry in the humming of the strings, there is music in the spacing of the spheres. -- Pythagoras ; ) . Always a pleasure. #CVPR2025 @CVPR.

RT @NateMankovich: I will be presenting A Flag Decomposition for Hierarchical Datasets at CVPR this afternoon. Come visit poster 282 to lea….

RT @trgokhale: The CVPR House Band returns! We're ready 😎.Show starts at 7pm in Grand Ballroom. Hope to see you all there!. @CVPR .#CVPR202….

Starting soon: my talk at the #C3DV Workshop of @CVPR on 3Cs of AI: .Compression, Composition, Cognition #CVPR2025.

Eylülde bekleriz. : ).

RT @t24comtr: 🧠 Foça’da bilim dolu bir ütopya: Bilimler Köyü. 📌 Bilimler Köyü, Foça’da, Foça Belediyesi’nin çok akıllıca bir iş yaparak bu….

RT @jacklangerman: 🚨🎉Today is the day! 🥳🚨. Room 104 D all day. Workshop on Urban Scene Modeling at @CVPR. kicking off at 9AM! . DONT MISS I….

RT @jacklangerman: @CVPR is right around the corner! Join us, June 11th for the 2nd Workshop on Urban Scene Modeling (@structGeom) for an i….

Another fruitful collaboration with @NateMankovich, being the third of a sequel.

We also provide ways to estimate flag types and show applications to few shot learning. We hope that our new tool makes it easier to make sense of hierarchies that can be explained by linear nested subspaces and make our implementation available: 👇.

We then propose the Flag Block Modified Gram-Schmidt (Flag-BMGS) algorithm for finding the flag decomposition (FD) and its robust version. Instead of orthonormalizing matrices to yield QR, Flag-BMGS operates on a matrix with a column hierarchy to produce a FD.👇

#FunWithFlags Episode III [. We introduce a new decomposition (like QR) but particularly suited to hierarchical datasets, where the hierarchy is to be respected. We use the natural ground offered by flag manifolds, nested sequences of subspaces. 👇 #CVPR

This is a first collaboration with the excellent Yiming Huang. We will soon release our implementation under: Stay tuned for more upcoming works under the common umbrella of topological generative models. See our CIRCLE Group:.

We show that our model exhibits stronger theoretical guarantees in generating realistic and high-fidelity graphs, confirmed by experimental evidence across real molecular and general graph generation benchmarks.👇

We represent the intermediary higher-order guides as cell complexes. A key step in our algorithm is the proposed cell complex filtering, which generates an intermediate state of a graph by pruning nodes and edges that do not belong to a given cell complex (see Figure).👇

We divide the fwd / reverse processes into K time windows, first generating coarse higher-order skeletons before refining them into pairwise relations. The steps are implemented by an OU diffusion bridge using Doob's h-transform to steer the SDE towards a specified endpoint. 👇

Instead of attempting to generate higher-order structures, we decompose the generation process into a composition of easy-to-learn, meaningful intermediate steps and utilize the higher-order cues during these intermediate stages of fwd / bwd processes. 👇

Relations beyond pairwise interactions are essential in networks. Yet, their role in semantic graph generative models remains unclear, as mentioned in our #ICML2024 paper [. We address this by integrating higher-order information in a principled manner.👇.