Publications

Conference Papers

Xie, S., Giraldo, J. H.
Variational Graph Contrastive Learning.
Accepted at NeurIPS 2024 Workshop: Self-Supervised Learning - Theory and Practice, 2024.
View Paper

Xie, S., Liu, Y. ,Shuai W. FTUnet: Feature Transferred U-Net for Single HDR Image Reconstruction.
Accepted at ACM Multimedia Asia (MMA), Oral Presentation, 2023.
View Paper

Xie, S., Zhu S. Feasibility Study of Intelligent Healthcare Based on Digital Twin and Data Mining.
Accepted at CISAI (International Conference on Computer Information Science and Artificial Intelligence), 2021.
View Paper

Patents

Xie, S., Yuan, R., Rossi, S., Hannagan, T.
Neural Network Parameter Diffusion.

• Patented method utilizing autoencoders and latent diffusion models to compress and generate experts in mixture-of-experts models.

Xie, S., Yuan, R., Rossi, S., Hannagan, T.
Permutation Symmetries Applied to DeepSeek Mixture of Experts Language Models.

• Developed a novel permutation symmetry method for aligning weights in GLU-based MoE architectures, ensuring efficiency in model compression, fine-tuning, and merging.
• Applied to DeepSeek-MoE-16B and Qwen1.5-MoE-A2.7B, demonstrating identical perplexity on the WikiText dataset and preserving generation consistency.

Reviewer

• Reviewer for NeurIPS 2024 Workshop on Compression.