Three pioneering studies by UC Riverside computer scientists will presented in July at the 42nd International Conference on Machine Learning (ICML) in Vancouver, Canada, showcasing research that pushes the boundaries of artificial intelligence in safety, privacy, and adaptability.
The papers offer vital advances in making AI safer to use, better at respecting privacy regulations, and more capable of adapting to constantly changing data environments. Each study was led by UCR researchers and selected through a highly competitive peer-review process, underscoring the university’s rising prominence in machine learning research.
“These studies tackle some of the most pressing challenges in AI today—from harmful output risks in vision-language systems to ensuring data deletion compliance and enabling learning in rapidly evolving environments,” said Amit Roy-Chowdhury, a co-author on two of the three studies and a professor of electrical and computer engineering at UCR.
Highlights of the three papers:
- Layer-wise Alignment: Examining Safety Alignment Across Image Encoder Layers in Vision Language Models
Lead Authors: grad students Saketh Bachu and Erfan Shayegani
This study reveals a previously unknown vulnerability in vision-language models, which combine images and text to answer questions. The researchers discovered that exiting early from the image processing layers—rather than using the full model—can cause these systems to produce harmful responses, even when images are safe. They term this vulnerability “Image enCoder Early-exiT” (ICET). To mitigate the risk, they introduce a new reinforcement learning method called Layer-wise Clip-PPO (L-PPO), which improves safety alignment across layers.
- A Certified Unlearning Approach without Access to Source Data
Lead Author: Umit Yigit Basaran
Addressing rising concerns over data privacy and new regulations, this paper proposes a novel “certified unlearning” framework that enables the erasure of specific data from machine learning models—even when the original training data is no longer available. The technique uses a surrogate dataset with similar statistical properties to simulate the data removal process and includes formal mathematical guarantees that the unlearned model mimics one retrained without the deleted data. This allows organizations to meet data deletion requests without retraining models from scratch.
- TMetaNet: Topological Meta-Learning Framework for Dynamic Link Prediction
Lead Author: Hao Li
Dynamic graphs—such as social networks or transaction records—are constantly changing, making them hard to model. TMetaNet introduces a meta-learning strategy that incorporates high-order topological features to help AI systems better predict new connections in evolving networks. It uses a novel representation method, Dowker Zigzag Persistence, to capture deep structural shifts over time, significantly boosting model robustness and accuracy. Real-world tests showed up to a 74.7% improvement in predictive performance over current methods.
Each of these works will be presented at ICML 2025, a prestigious venue for machine learning research globally. With hundreds of submissions, only a fraction of studies makes it to the conference’s main program, making the UCR trio’s selection especially notable.
“Our goal is to develop AI that not only performs well but behaves responsibly in real-world applications,” said Roy-Chowdhury. “These papers reflect that commitment—from aligning model safety to respecting user privacy and improving learning in complex, dynamic settings.”
These innovations further establish UC Riverside as a growing hub for high-impact AI research, with faculty and students leading work at the intersection of machine learning, ethics, and societal needs.