Research
I am broadly interested in the interplay between geometry and deep learning:
1) Deep learning for geometry:
My work focuses on neural methods for shape analysis and deformation. I'm actively pursuing two projects in this space; i) a new architecture for learning on surfaces, and ii) a shape descriptor that is robust to degenerate, multi-component geometry.
I am particularly interested in developing robust and efficient architectures suitable for large scale, in-the-wild 3D datasets.
2) Geometry for deep learning:
Using differential geometry to understand and control neural networks is particularly exciting. For example, augmenting inference using geometric ideas (e.g. as-smooth-as-possible latent interpolation), or understanding the geometry of high-dimensional feature spaces and how to exploit it for "editing" pretrained models.
I aim to extend such geometric ideas to large (vision/language/diffusion) foundation models, as the boundaries between domains continue to blur.
|
|
One Noise to Rule Them All: Learning a Unified Model of Spatially-Varying Noise Patterns
Arman Maesumi, Dylan Hu, Krishi Saripalli, Vladimir G. Kim, Matthew Fisher, Sören Pirk, Daniel Ritchie
ACM Transactions on Graphics (Proceedings of SIGGRAPH), 2024
[ project page / pdf (34mb) / arXiv / code / bibtex ]
A data augmentation strategy that enables diffusion models to smoothly interpolate between disjoint data modes. We train a diffusion model to blend multiple types of procedural noise patterns, even in the absence of "in-between" training data.
|
|
Explorable Mesh Deformation Subspaces from Unstructured 3D Generative Models
Arman Maesumi, Paul Guerrero, Vladimir G. Kim, Matthew Fisher, Siddhartha Chaudhuri, Noam Aigerman, Daniel Ritchie
SIGGRAPH Asia, 2023 (Conference Track)
[ project page / pdf (30mb) / arXiv / bibtex ]
What is the smoothest subspace that spans a set of points in latent space? We optimize smooth parametrizations of such subspaces in 3D generative models and use them to explore continuous variations of meshes.
|
|
Triangle Inscribed-Triangle Picking
Arman Maesumi
The College Mathematics Journal, 2019
[ pdf (1mb) / journal / bibtex ]
The probability density function and moments (OEIS A279055) of the area of stochastically generated inscribed geometry are derived.
Preliminary findings were presented at TUMC 2017.
|
Renderings
In my free time, I enjoy creating 3D renderings and physical simulations using various software. More can be found here. The programs and tools that I use include: Blender, Cinema 4D, RealFlow, Vray, Octane, Arnold, Krakatoa, and more.
|
Click to see more
|