Cognitive AI · 3D vision · Interpretability · Lifelong learning
My research bridges cognitive AI and NeuroAI with 3D vision, computer graphics, continual learning, interpretability, and reasoning. I study structured, compositional understanding of the world and the models we build, aiming to make artificial visual and perceptual intelligence more human-like. I am a PhD candidate in Computer Science at Johns Hopkins University, advised by Alan Yuille.
My work advances 3D part-based representation learning, bringing machine perception closer to how humans understand objects through their parts and relationships. This perspective connects my work on efficient 3D representations, unsupervised domain adaptation for 3D pose, joint part segmentation and naming, and scene understanding. My graphics work extends to editable animal fur reconstruction.
In continual learning, my contributions include formalizing catastrophic remembering and developing Share for almost strict continual learning. My work on adapter recycling explores how learned structure can be reused for efficient adaptation and inference.
In interpretable AI and the physics of AI, my work exposes stark failures of feedforward 3D reconstruction models, including hallucinated geometry from unrelated images or random noise. The Universal Weight Subspace Hypothesis provides evidence of shared low-dimensional weight structure across independently trained networks, with implications for efficient learning, inference, and continual adaptation.
Much of this work is first-authored or co-first-authored. Independent thinking, novelty, creativity, and ethical research guide my work.
I welcome researchers interested in connecting cognitive principles and the physics of learning with practical advances in 3D vision, graphics, and continual learning. Get in touch.
2026 · Efficient 3D reconstruction
Hexagonal surface elements with spatially varying color represent scenes with fewer primitives and lower memory requirements.
2026 · 3D animal reconstruction
Reconstructs editable, animatable animal fur from multiview images using explicit strands and a compact prior learned from human hair.
2025 · Physics of AI / explainable AI
Evidence that independently trained neural networks converge to shared, low-dimensional weight subspaces.
2026 · 3D vision / part semantics
Connects 3D part structure with semantic names, advancing part-based object understanding from a single example.
2026 · 3D evaluation / hallucination
Shows that learned 3D backbones can hallucinate dense geometry for unrelated scenes, repeated views, and random noise; introduces failure-aware consistency metrics.
2025 · efficient AI
Adapter recycling for resource-efficient adaptation and inference across LLMs, VLMs, diffusion, and 3D models.
Honeycomb-Surfels released: compact hexagonal primitives for efficient rendering and reconstruction.
FurE released: editable 3D animal fur reconstruction without animal-fur datasets.
Research talk at Apple with the multimodal AI team.
Research talk at Apple on Universal Weight Subspace, EigenLoRAx, and Name That Part.
Research talk at Brown's Serre Lab: Toward Human-Like Visual Intelligence: From 3D Parts to Universal Weight Subspaces.
Human-like 3D understanding through objects, named parts, spatial relationships, and compositional reasoning.
Remembering and forgetting, reusable representations, and efficient adaptation as tasks and environments change.
Interpretable representations, shared weight geometry, and principles of learning that explain model behavior and support robustness.