Hung Nguyen’s Personal Page
General Information
My name is Hung Nguyen, westernized from my Vietnamese name ‘‘Nguyễn Ngọc Duy Hưng’’. I recently graduated with my Master’s degree in Electrical Engineering at the University of California San Diego, United States, through a fully-funded scholarship (~$110,000 USD) from Vingroup. I am a graduate student researcher in Signal & Image Processing at the Video Processing Lab, supervised by Prof. Truong Nguyen. I work on AI-based 3D reconstruction and medical image processing, guided by a strong signal processing perspective. Before that, I graduated with a Bachelor’s degree in Mechatronics Engineering at the Ho Chi Minh City University of Technology, Vietnam, working on perception for robotics.
Research Interests
My research interests include, but not limited to, the following:
- How can digital signal processing (DSP) concepts (e.g., wavelets, Fourier transform) be incorporated in machine learning frameworks to enhance sparsity, generalizability and interpretability? The key idea is that many input forms (e.g., images, videos, texts) can be interpreted as signals, and hence subjected to processing and analysis under a DSP perspective. My belief is that without proper signal-based constraints, we risk letting the model learn representations that contribute less meaningfully to the task, which can have adverse effects on robustness and efficiency.
- How can different kinds of priors (e.g., task-specific, semantics) be leveraged to improve machine learning frameworks? As a designer of AI systems, I believe it is essential to understand domain knowledge of the task at hand and incorporate meaningful task-specific priors. Similar to signal processing, these task-specific priors can help improve robustness and efficiency. Besides, I am interested in incorporating semantics priors into machine learning frameworks, as they tend to be highly robust and align naturally with how humans perceive and reason about the world.
- How can these priors constrain the learnt representations under limited data? With a sparse set of labelled inputs, the learnt representations tend to be inadequately controlled, leading to subpar performances. However, working in low-data regime provides the benefits of shorter training time and cheaper annotation costs. The key idea is that the priors above, coupled with DSP concepts, can guide the model toward learning compact and structured representations that enhance robustness and efficiency, despite relying only on sparse inputs.
News
[06/2025] A paper on Learnable DWT for 3DGS optimization is accepted to the 2025 ICCV Workshop.
[06/2025] I successfully defended my Master’s thesis!
[05/2025] A paper on Learnable DWT with biorthogonal wavelets is accepted to the 2025 EUSIPCO conference.
[05/2025] A paper on Learnable DWT for CT image classification is accepted to the 2025 EUSIPCO conference.
[04/2025] I passed the Comprehensive Exam requirements and will graduate from UCSD with a Master’s degree! I will still be doing my Master’s thesis at UCSD for additional research experience.
Contact
Personal Email: nnduyhung3112 at gmail dot com
Institutional Email: hun004 at ucsd dot edu