April 24, 2025 - AI, Machine Learning and Computer Vision Meetup

This is a virtual event.

Register for the Zoom

Towards a Multimodal AI Agent that Can See, Talk and Act

The development of multimodal AI agents marks a pivotal step toward creating systems capable of understanding, reasoning, and interacting with the world in human-like ways. Building such agents requires models that not only comprehend multi-sensory observations but also act adaptively to achieve goals within their environments. In this talk, I will present my research journey toward this grand goal across three key dimensions.

First, I will explore how to bridge the gap between core vision understanding and multimodal learning through unified frameworks at various granularities. Next, I will discuss connecting vision-language models with large language models (LLMs) to create intelligent conversational systems. Finally, I will delve into recent advancements that extend multimodal LLMs into vision-language-action models, forming the foundation for general-purpose robotics policies. To conclude, I will highlight ongoing efforts to develop agentic systems that integrate perception with action, enabling them to not only understand observations but also take meaningful actions in a single system.

Together, these lead to an aspiration of building the next generation of multimodal AI agents capable of seeing, talking, and acting across diverse scenarios in both digital and physical worlds.

About the Speaker

Jianwei Yang is a Principal Researcher at Microsoft Research (MSR), Redmond. His research focuses on the intersection of vision and multimodal learning, with an emphasis on bridging core vision tasks with language, building general-purpose and promptable multimodal models, and enabling these models to take meaningful actions in both virtual and physical environments.

ConceptAttention: Interpreting the Representations of Diffusion Transformers

Recently, diffusion transformers have taken over as the state-of-the-art model class for both image and video generation. However, similar to many existing deep learning architectures, their high-dimensional hidden representations are difficult to understand and interpret. This lack of interpretability is a barrier to their controllability and safe deployment.

We introduce ConceptAttention, an approach to interpreting the representations of diffusion transformers. Our method allows users to create rich saliency maps depicting the location and intensity of textual concepts. Our approach exposes how a diffusion model “sees” a generated image and notably requires no additional training. ConceptAttention improves upon widely used approaches like cross attention maps for isolating the location of visual concepts and even generalizes to real world (not just generated) images and video generation models!

Our work serves to improve the community’s understanding of how diffusion models represent data and has numerous potential applications, like image editing.

About the Speaker

Alec Helbling is a PhD student at Georgia Tech. His research focuses on improving the interpretability and controllability of generative models, particularly for image generation. His research is more application focused, and he has have interned at a variety of industrial research labs like Adobe Firefly, IBM Research, and NASA Jet Propulsion Lab. He also has a passion for creating explanatory videos of interesting machine learning and mathematical concepts.

RelationField: Relate Anything in Radiance Fields

Neural radiance fields recently emerged as a 3D scene representation extended by distilling open-vocabulary features from vision-language models. Current methods focus on object-centric tasks, leaving semantic relationships largely unexplored. We propose RelationField, the first method extracting inter-object relationships directly from neural radiance fields using pairs of rays for implicit relationship queries. RelationField distills relationship knowledge from multi-modal LLMs. Evaluated on open-vocabulary 3D scene graph generation and relationship-guided instance segmentation, RelationField achieves state-of-the-art performance.

About the Speaker

Sebastian Koch is a PhD student at Ulm University and Bosch Center for Artificial Intelligence. He is supervised by Timo Ropinski from Ulm University. His main research interest lies at the intersection of computer vision and robotics. The goal of his PhD is to develop 3D scene representations of the real world that are valuable for robots to navigate and solve tasks within their environment

RGB-X Model Development: Exploring Four Channel ML Workflows

Machine Learning is rapidly becoming multimodal. With many models in Computer Vision expanding to areas like vision and 3D, one area that has also quietly been advancing rapidly is RGB-X data, such as infrared, depth, or normals. In this talk we will cover some of the leading models in this exploding field of Visual AI and show some best practices on how to work with these complex data formats!

About the Speaker

Daniel Gural is a seasoned Machine Learning Evangelist with a strong passion for empowering Data Scientists and ML Engineers to unlock the full potential of their data. Currently serving as a valuable member of Voxel51, he takes a leading role in efforts to bridge the gap between practitioners and the necessary tools, enabling them to achieve exceptional outcomes. Daniel’s extensive experience in teaching and developing within the ML field has fueled his commitment to democratizing high-quality AI workflows for a wider audience.