Back
On November 10, the Max Planck Institute for Intelligent Systems in Tübingen will host a special symposium on the future directions in intelligent systems. Keynote lectures by leading international scientists will offer insights into groundbreaking research and emerging trends, giving attendees a forward-looking perspective on the global development of intelligent systems.
Location: N0.002 Lecture Hall, MPI-IS Tübingen
09:15 - 09:30
15min
09:30 - 10:30
1h
Geometry First: Structured Representations for Generating 3D in a Large-Data World
Recent advances in machine learning have shown remarkable progress in generative synthesis of text, images, and even video, fueled by very large-scale data and compute. 3D, however, stands at the frontier of enabling machines to perceive and imagine the physical world -- critical for applications spanning content creation, mixed reality, and robotics. However, synthesis of 3D geometry remains constrained by significantly more limited data representing higher-dimensional information. In this talk, I will present our approach towards briging this gap by incorporating structured priors into generative 3D geometric synthesis. I will first describe how spatially grounded 3D scene representations can anchor learning, though often at a high-dimensional cost. To address this, I will introduce methods for constructing compact 3D representations that preserve essential structures while remaining highly expressive. Finally, I will show how integrating 3D structured priors with powerful complementary 2D signal can enable new possibilities in generative 3D synthesis. By integrating structured 3D priors, this opens the door to new paradigms beyond data limits for 3D synthesis.
Angela Dai is an Associate Professor at the Technical University of Munich where she leads the 3D AI Lab. Angela's research focuses on understanding how real-world 3D scenes around us can be modeled and semantically understood. Previously, she received her PhD in computer science from Stanford in 2018, advised by Pat Hanrahan, and her BSE in computer science from Princeton in 2013. Her research has been recognized through an ECVA Young Researcher Award, ERC Starting Grant, Eurographics Young Researcher Award, German Pattern Recognition Award, Google Research Scholar Award, and an ACM SIGGRAPH Outstanding Doctoral Dissertation Honorable Mention.
10:30 - 11:00
30min
11:00 - 12:00
Phi- Building Small Models with Frontier Capabilities
The Phi series of small open-weights models (1B-14B parameters) achieve best in-class performance on wide range of general-purpose language modeling tasks, often competitive with or outperforming 5-10x larger models. Due to the smaller size and the resulting capacity limits, these models generally tend to excel in reasoning heavy domains compared to knowledge heavy tasks. The key ingredient in the success of these models lies in diligent data and curriculum curation especially using frontier models and agents for synthetically generated content. In the latest model in the series, Phi-4-reasoning, we build on these strengths along with targeted post-training to enable the models to perform more complex non-linear reasoning using a form of "scratch pad'' designated for exploring multi-step, decomposition, internal reflection, and exploration of multiple problem-solving strategies. In this talk I will focus on high-level methodology and insights that went into building these models.
Dr. Suriya Gunasekar is a Principal Research Manager at Microsoft Research. Her current work is on improving language model capabilities through scale, data curation, and creative uses of synthetic data generation, while maintaining broader interests in evaluation and alignment of AI systems for targeted use-cases. In the recent past, she has worked in areas adjacent to science of deep learning and theoretical ML. Prior to MSR, she was a Research Assistant Professor at Toyota Technological Institute at Chicago.
Registration Closed
This event is no longer open for registration.
If you have any questions, feel free to contact us:
More information