Projects

Research projects on data-efficient foundation models, multimodal perception, and federated learning.

Data-Efficient Foundation Models & Multimodal Systems

I work on making large models more practical under real constraints:

  • Foundation Models & Data Selection — select small, high-impact subsets for LLMs and vision models using training dynamics and attention structure.
  • Multimodal Perception & Robotics — build lightweight, low-latency perception systems using event-based sensing and efficient architectures.
  • Federated & Distributed Learning — reduce communication, preserve privacy, and keep performance competitive in real-world, heterogeneous settings.

Foundation Models & Data Selection

Select less, learn more: Develop methods like attention-based token saliency and loss-trajectory coresets to train LLMs and vision models efficiently without sacrificing accuracy.

project thumbnail

TRIM

Token-wise attention–derived saliency for efficient instruction-tuning data selection (forward-only).
project thumbnail

CLD

Coresets via correlation of loss differences (training-dynamics selection with theory).

Multimodal Perception & Robotics

Perceive faster, on cheaper hardware: Use event cameras and compact neural architectures to enable real-time detection and tracking for robotics and embedded platforms.

project thumbnail

DOTIE

Detecting Objects through Temporal Isolation of Events

Federated & Distributed Learning

Learn collaboratively, communicate less: Build communication- and data-efficient federated learning methods that handle heterogeneous clients while preserving privacy and model quality.

project thumbnail

TOFU

Toward Obfuscated Federated Updates — Encoding weight updates into gradients from proxy data for communication-efficient and privacy-enhanced federated learning.