Research Area

Infographic showing the progression from visual input to transferable features and robust adaptation.

Abstract

Computer Vision and Learning Algorithms

We study robust vision and learning algorithms for recognition, representation, and adaptation at scale. This area investigates scalable visual representations and adaptation strategies that remain reliable across data shifts, label scarcity, and long-tail conditions. Foundational vision and learning methods for robust generalization

Visual RecognitionLarge-scale ModelMeta Learning
Core Architecture figure

Core Architecture

Designing efficient SOTA architectures with manageable computational cost

Representation Learning figure

Representation Learning

Reinforcing feature representation through token pooling, hashing, and continual learning

Vision-Language Model figure

Vision-Language Model

Adapting large VLM to new tasks and domains through parameter-efficient transfer learning

Infographic showing large-model compression, runtime optimization, and practical language-model deployment.

Abstract

Efficient Learning for LLMs

We develop practical methods for making large language models lighter, faster, and more efficient to deploy. This area focuses on making large language and multimodal models lighter, faster, and easier to deploy while preserving reasoning quality. Efficient and practical multimodal intelligence for deployment

Model CompressionEfficient LearningSystem-level Optimization
Pruning figure

Pruning

Removing redundant parameters and structures from large models while preserving performance

Quantization figure

Quantization

Reducing numerical precision of weights and activations to lower memory and compute cost

Efficient Transfer Learning figure

Efficient Transfer Learning

Adapting pretrained models to new tasks and domains with minimal additional training cost

Infographic showing a robot learning pipeline from 3D-aware perception through vision-language reasoning to action.

Abstract

Robot Learning

We build robot learning pipelines that connect visual understanding to efficient and reliable robotic behavior. This area connects visual-language understanding to robot policies for practical perception-to-action pipelines in real environments. Embodied AI from perception to reliable robot behavior

VLAEfficient Robotics3D Recognition
Multimodal World Model figure

Multimodal World Model

Learning world dynamics to understand causality and dynamics in physical environments

Robust & Generalizable VLA figure

Robust & Generalizable VLA

Improving robustness and generalization across diverse tasks, environments, and embodiments

Speech & Sound VLA figure

Speech & Sound VLA

Grounding speech instructions and sound for reactive robot control instead of relying on text

Infographic connecting industrial fault analytics, medical image intelligence, and deployment validation.

Abstract

Industrial and Medical AI

We apply AI methods to industrial reliability and medical analysis, with a focus on high-impact real-world use cases. This area applies machine learning to industry and healthcare with emphasis on reliability, interpretability, and robust validation in real use cases. Trustworthy AI for industrial reliability and medical analysis

Battery AIFault DetectionMedical AI
Medical Image Diagnosis figure

Medical Image Diagnosis

Building diagnostic models robust to long-tail distributions and domain shift problem

Real-World Visual Tracking figure

Real-World Visual Tracking

Applying vision systems to tracking, recognition, and defect detection in real-world environments

Battery Prognostics figure

Battery Prognostics

Predicting battery degradation and remaining useful life by fusing heterogeneous signals

Browse all publications