Projects
MT3D
A Gaussian splatting based 2D lifting technique that generates 3D representation by learning the basic shape/structure from a high-fidelity 3D object.
LinkPolyINR
Developed Poly-INR, the first INR-based model that can represent complex shapes and generate large, diverse datasets like ImageNet. Models images as polynomial functions of their coordinates, progressively increasing polynomial degrees through multiple MLP layers.
RNAS-CL
Proposes Robust Neural Architecture Search by Cross-Layer Knowledge Distillation (RNAS-CL), optimizing accuracy, latency, and robustness without adversarial training. Demonstrates student models inheriting robustness from adversarially robust teacher models.
LinkAdjoined Network
Introduces Adjoined Networks (AN), a novel training approach that simultaneously compresses and regularizes CNN-based architectures. AN trains the base and compressed networks together, with the compressed network's parameters being a subset of the base network's.
Link