Join our virtual meetup to hear talks from experts on cutting-edge topics across AI, ML, and computer vision.

In an era where model complexity and deployment constraints increasingly collide, achieving both scalability and efficiency in deep learning has become essential. Scalable and efficient deep learning ensures that powerful models can be trained, deployed, and adapted under limited computational and data resources, enabling broader accessibility and practical application. From understanding to generation, this talk unifies methods that cut costs while preserving capability.

Generative video models have made rapid progress in visual realism, yet they frequently violate basic physical laws, producing implausible motion and incorrect cause-effect relationships. This talk presents MoReGen, a physics-grounded, agentic text-to-video generation framework that integrates Newtonian physics directly into the generation process via executable physics-engine code. By coupling vision–language models with trajectory-based physical evaluation and iterative feedback, MoReGen produces videos that are both visually coherent and physically consistent. We further introduce MoRe Metrics and MoReSet, a benchmark and dataset designed to evaluate physics fidelity beyond appearance-based metrics such as FID and FVD. Together, this work demonstrates a path toward visual AI systems that reason about motion, interaction, and causality in the real world rather than hallucinating them.

World models have emerged as a foundational paradigm for enabling agents to simulate, predict, and reason about complex environments. Recent advances driven by diffusion transformer (DiT) architectures have dramatically expanded the fidelity, scalability, and physical plausibility of learned world models. In this work, we present a world model framework built upon the diffusion transformer paradigm, following the design philosophy of state-of-the-art systems such as NVIDIA Cosmos. Our approach comprises three core components: (1) a spatiotemporal variational autoencoder (VAE) that compresses high-resolution video into a compact continuous latent space with strong temporal causality, enabling efficient encoding and decoding of long-horizon video sequences; (2) a transformer-based diffusion backbone that operates on 3D-patchified latent tokens, leveraging self-attention and cross-attention with text embeddings to iteratively denoise Gaussian noise into physically coherent future video states using a flow matching objective; and (3) a scalable pre-training and post-training pipeline that first trains a generalist world foundation model on large-scale, diverse video data and then specializes it to target physical AI domains — such as robotic manipulation, autonomous driving, or embodied navigation — through task-specific fine-tuning. Our model supports both text-to-world and video-to-world generation, enabling action-conditioned future state prediction for downstream planning and policy learning. We discuss implications for synthetic data generation, sim-to-real transfer, and the integration of world models into vision-language-action (VLA) pipelines for physical AI.

The Physical AI Research (PAIR) initiative advances the next frontier of artificial intelligence: enabling systems that can perceive, reason, and act reliably in the physical world. By uniting expertise across engineering, computer science, health sciences, and the social sciences, PAIR develops safe, transparent, and human-aligned AI that bridges digital models with real-world dynamics. The initiative is organized around three intellectual pillars: Learning and Modeling the World, through physics-informed multimodal learning, realistic simulations, and digital twins; Reasoning in the World, by integrating multimodal evidence to support grounded decision-making under uncertainty; and Acting in the World, by ensuring AI systems are verifiable, explainable, energy-efficient, and trustworthy. Together, these efforts position Physical AI as a foundational science driving innovation in health, sustainability, and security.