3D vision & cognitive AI

My research bridges cognitive AI and NeuroAI with 3D computer vision, computer graphics, continual learning, interpretability, and reasoning. I investigate structured, compositional understanding of both the world and learned models, with the goal of making artificial visual and perceptual intelligence more human-like.

A central focus is advancing 3D part-based representation learning: connecting geometry, semantic meaning, and relationships between parts and objects. This perspective informs my work on efficient 3D representations, unsupervised domain adaptation, joint part segmentation and understanding, scene understanding, and animal fur reconstruction. Alongside it, my work on catastrophic remembering, strict continual learning, and the physics of AI investigates how models learn, retain knowledge, and fail, with efficiency, robustness, and interpretability as central goals.

Analysis by synthesis is one aspect of this program: constructing and testing generative explanations of visual observations. Other directions study perceptual organization, remembering and forgetting, and shared structure across learned models.

Research directions

3D part-based understanding

Focus: Part-based representation learning for structured, compositional understanding of 3D objects.

Approach: Connect part geometry, semantic naming, and compositional representations to support human-like visual understanding.

Selected work: Name That Part, structured 3D compositional models, and synthetic-to-real part understanding.

Human-like perception & robust reasoning

Problem: Models recognize objects but fail to reason over hierarchy, attributes, and affordances.

Key idea: Connect cognitive principles of perceptual organization with interpretable representations and diagnostic evaluation.

Selected work: Perceptual Taxonomy and Can These Views Be One Scene?.

Interpretable AI & physics of AI

Focus: Understand the representations, weight geometry, and learning dynamics that shape model behavior.

Approach: Study shared low-dimensional structure across trained models to explain how learning and adaptation work.

Selected work: The Universal Weight Subspace Hypothesis.

Continual learning & efficient adaptation

Problem: Continual adaptation is compute-heavy and prone to forgetting.

Key idea: Shared low-dimensional update geometry for reuse, merging, and strict continual learning.

Selected work: EigenLoRAx, Share, and continual forgetting and remembering.

Selected Contributions and Research Ideas

  1. 3D part-based understanding. Advanced structured object understanding through named 3D parts, joint segmentation and semantic alignment, and compositional representations.
  2. Universal weight geometry / shared subspaces. Studied and characterized shared low-dimensional structure across trained neural models.
  3. Hierarchical perceptual diagnostics. Developed a framework for evaluating compositional and structured visual reasoning.
  4. Catastrophic remembering + strict continual learning. Formalized catastrophic remembering and developed shared-subspace methods for almost strict continual learning.
  5. Source-free, image-only UDA for 3D pose (3DUDA). Introduced a source-free adaptation setting for 3D object pose estimation.
  6. Adapter recycling / principal subspaces. Reused adapters to estimate principal subspaces for efficient adaptation and reuse.
  7. Shared subspace continual fine-tuning (Share). Framed continual updates as shared-subspace fine-tuning for efficiency and stability.

Research Roadmap

A living view of completed recent projects, active directions, and the longer-term research program.

Completed recent works

  • Completed: Strict continual learning methods for large language models.
  • Completed: Reliable evaluation metrics for 3D reasoning systems.
  • Completed: 3D animal fur reconstruction and shape modelling.

Next directions

  • Efficient 3D primitive representations.
  • Probabilistic and semi-supervised semantic 3D part segmentation.
  • Part-based vision-language-action models.
  • LLM/VLM model merging and transfer.

Long-term direction

  • General theory of deep learning and scaling behavior.
  • Synthetic internal 3D worlds for part-level causal and physics reasoning.
  • Explore super-Turing perspectives (theoretical).