Selected Research

Grounded & Predictive Intelligence

Preprint

Momentum-Guided Semantic Forecasting (MoFore)

A self-supervised video representation learning framework that predicts future semantic features across temporal gaps. Momentum-guided teacher supervision and contrastive regularization encourage motion-aware, temporally consistent representations without requiring action labels during training.

Click a slide to enlarge.

Efficient Learning & Optimization

Preprint

Memory-Efficient Contrastive Learning via Budgeted Hard Negative Selection

A memory-efficient contrastive learning framework combining blockwise computation, streaming hard-negative selection, and online pruning to reduce the memory demands of vision-language representation learning.

Key result. Scaled ViT-B/32 contrastive training to batch size 8,192, compared with a dense baseline that ran out of memory at 4,096 on the tested GPU.

Robust Model Evaluation & Physical Reasoning

Ongoing Research

Ongoing direction

Self-Learning & Predictive Intelligence

Exploring representations that learn from visual and physical dynamics, moving beyond static recognition toward predictive understanding.