Arman Maesumi

arman_maesumi@brown.edu

I am a rising fifth year Computer Science PhD student at Brown University, where I am advised by Professor Daniel Ritchie. My work is supported by the NSF Graduate Research Fellowship Program.

I received my BS in computer science from The University of Texas at Austin in 2021, where I did research with Professor Chandrajit Bajaj.

My research interests lie at the intersection of deep learning and computer graphics. In my free time I enjoy creating 3D art, cooking, and playing chess.

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.

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.

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.

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.

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.

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.

Template