Large Language Model (LLM) Unlearning
LLM unlearning methods fall into two groups: divergence-based approaches, which push models away from pretrained parameters to undo influence of a forget set, and convergence-based methods, which guide models toward fixed target behaviors. Both face core trade-offs between under-forgetting, over-forgetting, and preservation of general utility. Our work introduces principled frameworks for precise, targeted forgetting that preserves model coherence.
Large Vision-Language Model (LVLM) Spatial Reasoning
Spatial reasoning is an emerging capability of multimodal foundation models to understand and manipulate spatial relationships between objects and environments while grounding reasoning in natural language. Our research enables LVLMs to perform pixel-grounded, depth-aware, and spatially-aligned reasoning tasks essential for pedestrian navigation, robotics, accessibility, and embodied AI.
LLM Safety & Adversarial Robustness
LLM safety research spans alignment and preference optimization, adversarial robustness (jailbreaks, prompt injection), distributional shifts, content safety, and evaluation/benchmarking. Our work includes novel backdoor attack algorithms, adversarial in-context learning, and defenses against hijacking.
Explainable & Interpretable AI
Interpretable Machine Learning and Explainable AI (XAI) research addresses both developer-facing understanding (how information flows inside DNNs) and end-user-facing explanation (what a model "sees" in an image or "comprehends" in text). Original contributions span explainable model prediction and interpretable feature representation learning.
Algorithmic Fairness
Fairness in AI addresses biases embedded in training data and model design that lead to disparate outcomes across demographic groups. Our work develops methods that jointly promote fairness, explainability, and robustness in deep neural networks, including vision transformers.
AI for Health
Use-inspired AI research applies foundational methods to pressing human-centered challenges in medical imaging, clinical risk assessment, mobility equity, and accessibility. Work in this area has been featured in Detroit PBS, WJR Radio, and Hour Detroit Magazine.
AI for Mobility & Social Good
AI-driven solutions for mobility equity, accessibility, and urban infrastructure — including pedestrian navigation, EV charging optimization, and automated segmentation of mobility environments for people with disabilities.