Raymond Yeh
Purdue University, IN·Computer Science·Early-career
Modern computer vision systems can recognize, edit and generate images remarkably well. Yet they store what they learned from training data in ways that are hard to inspect, trace or correct.
- Fields
Computer Science
Engineering
- Topics
computer vision
machine learning
deep learning
- Grants
- 2 known grants, through Jul 2031
What they work on
- Develop semi-parametric vision architectures that consult training examples during prediction, so each output can be traced to specific data, combining deep network accuracy with k-nearest-neighbor transparency.
- Apply semi-parametric models to unlearning, data attribution and model customization. Improve how models use stored data at test time through equivariant architecture design and new training strategies.
- Develop methods to remove unwanted or harmful concepts, such as copied artistic styles or celebrity likenesses, from trained AI image generators.
- Immunize AI image generators against fine-tuning on harmful concepts using bi-level optimization, and design efficient solvers that exploit distributed computation and sparsity.
More about the research
That opacity makes it hard to remove harmful abilities, explain an output or update a model safely. This research makes vision models and AI image generators more controllable. It takes two routes: building models that keep explicit links to their training data, and developing ways to remove or block harmful content generation.
Why it matters
AI image tools are now widely available, so their misuse, including copying artists' work, fake celebrity images and privacy harms, affects artists, public figures and anyone who shares photos. Methods to remove or block harmful generation, and vision models whose outputs can be traced and corrected, could make deployed AI systems safer, more accountable and easier to maintain.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Dongxiao Zhu
Wayne State University, MI·Computer Science
Develop optimization algorithms that make large language model unlearning, removing specific data influences without retraining, more effective, robust and efficient.
Siwei Lyu
University at Buffalo, NY·Computer Science
Strengthen deepfake detectors against anti-forensic attacks by processing each input with a randomly chosen model from an infinite ensemble of classifiers, making detectors harder to fool.
Faez Ahmed
Massachusetts Institute of Technology, MA·Engineering·Early-career
Develop diffusion models, generative AI that builds outputs by removing noise step by step, that learn from the full history of design optimization runs rather than only final optimized designs, to generate more precise engineering designs.
The school
Purdue University: 43 labs on PI Pages, $932M a year on research, 833 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.