Check out this media Article about my recent work.
Human-like perception, structured 3D world models, and continual & efficient learning.
"I study how structured world models and reusable learning dynamics enable human-like, lifelong intelligence under real constraints."
- Human-like Perception (CogSci-inspired)
- Analysis by Synthesis & 3D World Modeling
- Efficient & Strict Continual Learning (Reusable Structure)
News
One co-first authored paper accepted at International Conference on Computer Vision (ICCV)! See you in Hawaii!
Gaussian Scenes has been accepted by Transactions on Machine Learning Research (TMLR)!
Two papers accepted as Oral presentations at CVPR Workshops! See you in Nashville!
EigenLoRAx is now available. Try it out for efficient and sustainable learning of large models.
Our work on Sparse View 3D reconstruction is now available. Check out Gaussian Scenes.
I build cognition-inspired systems using analysis by synthesis, 3D world modeling, and reusable learning dynamics under real compute and data constraints.
Featured Work
Universal Weight Subspace Hypothesis
A mechanistic account of adaptation geometry that turns fine-tuning into reusable structure.
Problem: We lack a mechanistic explanation for why small updates enable robust adaptation across tasks.
Key idea: Learned updates concentrate in shared low-dimensional subspaces that can be reused and merged.
Why it matters: It enables parameter-efficient adaptation, merging, and strict continual learning with reduced overhead.
Name That Part
Turns part understanding into a scalable semantic engine—without privileged compute or data.
Problem: 3D systems often separate geometric part segmentation from semantic naming.
Key idea: Jointly align parts and names as a set alignment problem with a scalable annotation engine.
Why it matters: It yields consistent, reusable part semantics under real supervision constraints.
Perceptual Taxonomy
Diagnoses where VLM perception breaks: hierarchy, attributes, affordances, and function.
Problem: Vision-language models succeed at recognition but fail at structured perceptual reasoning.
Key idea: A cognitively grounded taxonomy and evaluation that probes how models reason.
Why it matters: It guides structure-based perception models and interpretability diagnostics.
EigenLoRAx
Recycles prior adapters into a principal subspace for fast, resource-efficient adaptation.
Problem: Adapter-based fine-tuning is efficient but its gains are often siloed.
Key idea: Recycle adapters to estimate principal subspaces for fast reuse and merging.
Why it matters: It accelerates adaptation while keeping continual updates lightweight.
Visibility
Recent work has been widely discussed across research and practitioner communities, including community digests and social threads.