Nghia H. (H.) Nguyen - /ŋɪə ŋwɪn/

Nghia H. (H.) Nguyen - /ŋɪə ŋwɪn/

PhD student at the University of Pennsylvania. My first name, Nghia, is pronounced like 'near', but you replace 'n' with 'ng'.

Research interests

I am interested in leveraging optimization and algebraic structures (e.g., the linear structures of the semantic concepts in LLMs) for trustworthy machine learning (interpretability and adversarial robustness). The research agenda I would like to pursue next consists of the following two directions:

  • Utilizing algebraic structures to design new algorithms in interpretability: Given that we know certain structures exist in the latent space of LLMs (e.g., linear representations, hierarchical orthogonality), how can we design new algorithms that take into account this prior knowledge to recover semantic concepts that are more reliable and more faithful to the underlying concepts used by the model?
  • Understanding the emergence of these algebraic structures through optimization objective/dynamics: The fact that these algebraic structures exist is not random but a deep correspondence to certain independence conditions in the real world. As an example, certain semantic concepts are orthogonal in the latent space of LLMs because they are compositional in real life and thus are reflected in the training data (e.g., color and shape of an object). What are all the structures that exist and how can we prove their emergence through a simple softmax-based loss?

Updates

  • 2025-11: I have a new preprint titled “Hierarchical Concept Embedding & Pursuit for Interpretable Image Classification”! [Link]
  • 2025-09: I have a new paper under review titled “SSD: Sparse Semantic Defense against Semantic Adversarial Attacks to Image Classifiers”! [Link]
  • I’m starting a PhD journey in Fall 2023 at the Computer and Information Science Department, University of Pennsylvania, advised by René Vidal.
  • I’m excited to have a NeurIPS 2022 paper titled “Improving Neural Ordinary Differential Equations with Nesterov’s Accelerated Gradient Method”! TLDR: we introduce Nesterov accelerated gradients into Neural ODEs to make Neural ODEs more efficient without sacrificing accuracy. [Personal link] [OpenReview]

Education

University of Pennsylvania - Philadelphia, PA, US

  • PhD, Computer and Information Science (2023-present), Advisor: René Vidal.

University of Science, Vietnam National University - Ho Chi Minh City, Vietnam

Previous research experience

Previously, I was in the AI Residency program at FPT Software AI Center, one of the only two AI residency programs in Vietnam, working with Thieu N. Vo and Tan M. Nguyen.

For my undergraduate thesis (2021), I worked on the problem of Human-Object Interaction Detection. The thesis is on the topic of Visual-Language Reasoning, which interested me because I wanted to explore the synthesis of the profound progress made by the Computer Vision and Natural Language Processing communities.

In the summer of 2020, I was a Research Intern at the Distributed Information Systems Laboratory (LSIR), École Polytechnique Fédérale de Lausanne (EPFL). I worked on utilizing Graph Neural Networks for SQL Query Optimizers.

More about me

  • My full name (in the Vietnamese order) is Nguyen Ho Huu Nghia (Nguyễn Hồ Hữu Nghĩa - with diacritical marks).