I build adaptive learning systems that decide what information to process, when to compute, and how to generalize under real-world constraints.
I am an incoming Ph.D. student in Computer Science at the University of Maryland, College Park, where I will be advised by Dr. Ritwik Gupta. Before starting at UMD, I worked as a Research Engineer with Dr. Trevor Darrell’s group at UC Berkeley (BAIR) and as a Visiting Student Researcher in the NeuroAgents Lab at Carnegie Mellon University led by Dr. Aran Nayebi.
Currently
- Starting my Ph.D. at UMD, focused on adaptive perception and self-supervised systems.
- Working on efficient foundation models, long-sequence vision, remote sensing, and world-model based skill discovery.
- Open to collaborations on adaptive computation, robust perception, climate and disaster response, and embodied intelligence.
Selected Research
REOrdering Patches Improves Vision Models
Sequence models typically flatten images using a fixed row-major order. In long-sequence vision transformers, this ordering can affect performance because architectural approximations break full permutation invariance. We proposed REOrder, a two-stage framework for discovering task-optimal patch orderings using an information-theoretic prior and a Plackett-Luce policy optimized with REINFORCE.
| REOrder improves top-1 accuracy over row-major ordering by up to 3.01% on ImageNet-1K and 13.35% on Functional Map of the World. Project website | NeurIPS 2025 page |
Research Interests
My research focuses on self-supervised systems that allocate computation and representation adaptively for the real world. I work across computer vision, remote sensing, and reinforcement learning, with recent projects in physics-guided graph learning, long-sequence vision models, and world-model based skill discovery.
I am especially interested in attention and perception strategies that improve efficiency, robustness, and generalization for climate, disaster response, and embodied intelligence applications.
I previously completed my M.S. in Computer Science at the University of Pittsburgh, where I worked with Dr. Xiaowei Jia on machine learning for climate and environmental systems. My thesis examined the effectiveness of in-domain pretraining for remote sensing tasks.
Outside of research, I enjoy backpacking, film, music, and cats.
Feel free to contact me at declank @ umd.edu. I am open to new collaborations.