Aug 27 - AI, ML, and Computer Vision Meetup

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

Date, Time, and Location

Aug 27, 2026
9:00 AM - 11:00 AM PST
Online. Register for the Zoom!

Robust Concept Protection against Diffusion-Based Image Editing and Personalization

Diffusion-based image editing and personalization models have made it increasingly easy to manipulate and replicate visual concepts from only a few reference images. However, existing protection methods often overfit to a single attack model and fail to generalize across diverse editing pipelines.

In this presentation, I will discuss recent advances in concept protection for generative AI systems, focusing on targeted perturbation strategies and style-sensitive diffusion representations. I will also present experimental findings across multiple editing and fine-tuning scenarios, highlighting the challenges of robustness, transferability, and imperceptibility in practical protection settings. Finally, I will discuss open problems and future directions toward trustworthy generative content ownership.

About the Speaker

Qiuyu Tang is a Ph.D. student in Computer Science and Engineering at Lehigh University. Her research focuses on trustworthy AI, media forensics, and robust protection methods against diffusion-based image editing and personalization systems. Her recent work explores concept protection, style safeguarding, semantic image manipulation, and generative AI robustness. She has contributed to multiple publications in computer vision and AI safety, including research on diffusion model protection and manipulation detection, and has also served as a conference workshop organizer.

From Pixels to the Planet: Building Scalable and Grounded AI for Science

AI has demonstrated a lot of new possibilities, from drafting emails to image editing and generation. The efficacy of AI models is largely built upon a standard machine learning pipeline, where data is fed into models to get representations and predictions, and the performance is evaluated with controlled benchmarks and metrics. However, the mismatch arises when we try to transit this pipeline to the interaction with the real world and use AI for scientific discovery. Beyond close-set decisions, scientists want to discover new categories and propose new hypotheses. In this talk, I will share how I address the challenges of AI for science from the perspectives of data-centric methods and interpretability approaches.

About the Speaker

Jianyang Gu is a postdoctoral scholar at The Ohio State University. His research focuses on using data-centric methods to build scalable and interpretable foundation models for science.

Building Real-World Computer Vision Systems with Voxel51

This talk will explore practical workflows for building, evaluating, and improving modern computer vision systems. We’ll dive into real-world approaches to dataset curation, model analysis, multimodal AI workflows, and production-ready vision pipelines using open-source technologies.

The session is designed for engineers, researchers, and AI practitioners looking to better understand how teams are developing and scaling computer vision applications today. Expect practical demos, technical insights, and discussions around the evolving AI tooling ecosystem.

About the Speaker

Daniel Gural is an expert in Physical AI and has been working in the field for over 8 years. Working across healthcare he has experience in both operating use case as well as using Visual AI as an aid in psychology applications as well.

Seeing Is Not Enough: Visual Grounding, World Models and Why Computer-Use Agents Fail at Step 17

Computer-use agents are a computer vision problem wearing an agent costume: the model has to read a screen it has never seen, locate the right control among hundreds of near-identical ones, and understand what changed after it clicked — thirty times in a row without drifting. This talk covers the two things we built at Obin AI to make that work: a video-based world model that learns the latent dynamics of GUI and document environments, so the agent can roll out a candidate workflow internally and pick the branch that doesn't dead-end, and a checkpoint-based evaluation harness that scores verifiable intermediate states rather than required action sequences.

I'll be specific about the vision side — what the model needs to represent about an interface and what it can safely ignore, why optimising for target-state reachability beat optimising for reconstruction fidelity, and how we separate visual grounding failures from planning failures, since they look identical in the logs and need completely different fixes. The arithmetic that motivates all of it: at 90% per-step reliability a thirty-step workflow completes 4% of the time, which is why per-step benchmarks flatter these systems so badly.

I'll also cover what didn't work — deep rollouts against an imperfect dynamics model, and step-level preference data that produced more plausible-looking actions and fewer completed tasks.

About the Speaker

Nevasini Sasikumar is a machine learning researcher at Obin AI, where she builds video-based world models and computer-use agents that read screens and operate desktop and web software autonomously, with finance back-office automation as the deployment domain.