This project aimed to develop a robust face recognition system for occluded faces. My contributions included designing the core adaptive feature selection module and implementing the data preprocessing pipeline.
Published in International Conference on Artificial Neural Networks (ICANN), 2025, 2024
This work introduces HCNQA, a method to enhance spatial reasoning in 3D-VQA. By integrating a Hierarchical Concentration Narrowing (HCN) module, we guide the model’s attention to suppress shortcuts and improve performance on the ScanQA benchmark.
Investigated and implemented a pipeline to fuse textual spatial embeddings from a 3D grounding model (EDA) into an LLM (LEO), providing critical insights into mitigating spatial information loss in multimodal models.
Leading a project to develop a Neural Module Network (NMN) for 3D-VQA. My work involves designing a framework to parse natural language questions into executable programs, aiming to improve model interpretability, structured reasoning, and generalization.
Upcoming research on more variable and expressive latent representations of physical properties, moving beyond deterministic point estimates to better capture material uncertainty inherent in visual data.