Research

Efficient Machine Learning
As machine learning models—especially large language models—grow in size and complexity, so do their computational and environmental costs. My research in efficient machine learning is motivated by the need to make these powerful models more practical, sustainable, and accessible. I explore strategies such as model compression, pruning, and low-rank approximations to reduce resource requirements while preserving performance. The overarching goal is to design scalable learning systems that retain their capabilities across a wide range of devices and deployment settings.

Selected Publications:

• Lu Sun and Jun Sakuma, “Learning Semi-Structured Sparsity for LLMs via Shared and Context-Aware Hypernetwork”, in Proceedings of The 14th International Conference on Learning Representations (ICLR 2026), 2026, Rio de Janeiro, Brazil.
  
• Ke Bian, Lu Sun and Dengji Zhao, “Learning Compact Neural Networks via Generalized Structured Sparsity”, in Proceedings of the 27th European Conference on Artificial Intelligence (ECAI 2024), accepted, 2024, Santiago de Compostela, Spain.
  
• Tianxiao Cao, Lu Sun, Canh Hao Nguyen and Hiroshi Mamitsuka, “Learning Low-Rank Tensor Cores with Probabilistic ℓ0-Regularized Rank Selection for Model Compression”, in Proceedings of the 33rd International Joint Conference on Artificial Intelligence (IJCAI 2024), accepted, 2024, Jeju, Korea.
  
• Jiahui Xu, Lu Sun and Dengji Zhao, “MoME: Mixture-of-Masked-Experts for Efficient Multi-Task Recommendation”, in Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2024), 2527-2531, 2024, Washington D.C., USA.
  
  

Interpretable Machine Learning
In high-stakes domains like healthcare, law, and finance, understanding why a machine learning model makes a certain prediction is just as important as the prediction itself. My research in interpretable machine learning is driven by the challenge of making complex models more transparent and trustworthy. I am particularly interested in methods that can identify and explain which features are most relevant for individual predictions, enabling personalized and context-aware insights. This line of work aims to bridge the gap between model performance and human understanding, empowering users to make informed decisions based on model outputs.

Selected Publications:

• Lu Sun and Jun Sakuma, “Learning Local Feature Masks with Variational Information Bottleneck”, in Proceedings of the 35th International Joint Conference on Artificial Intelligence (IJCAI-ECAI 2026), 2026, Bremen, Germany. (To appear)
  
• Luhuan Fei, Weijia Lin, Jiankun Wang, Lu Sun, Mineichi Kudo and Keigo Kimura, Multi-View Multi-Label Personalized Classification via Generalized Exclusive Sparse Tensor Factorization. Knowledge and Information Systems, accepted, 2025.
  
• Jiankun Wang, Luhuan Fei and Lu Sun, Multi-level network Lasso for multi-task personalized learning, Pattern Recognition, 2024.
  
• Weijia Lin, Jiankun Wang, Lu Sun, Mineichi Kudo and Keigo Kimura, “Multi-Label Personalized Classification via Exclusive Sparse Tensor Factorization”, in Proceedings of the 23rd IEEE International Conference on Data Mining (ICDM 2023), 2023, Shanghai, China.
  
  

AI for Science and Materials Discovery
My research in AI for science focuses on developing generative and geometric machine learning methods for materials discovery, particularly for crystalline materials. I am interested in building efficient generative models that can rapidly propose physically meaningful and experimentally realizable crystal structures while respecting the geometric and symmetry constraints of material systems. My recent work explores few-step and one-step generative modeling on non-Euclidean crystal manifolds, combining flow matching, geometric deep learning, and transport-based generative modeling to significantly reduce the computational cost of crystal generation. I am also interested in synthesizability-aware generative modeling, aiming to bridge the gap between computationally stable materials and materials that can be practically synthesized in laboratories.