Shifeng XIE bio photo

About Me

Here is Shifeng XIE.
I am a PhD student in Prof. Themis Palpanas’s group at Université Paris Cité. Previously, I studied engineering at Télécom Paris and the Institut Polytechnique de Paris. My research interests focus on time series foundation models and agentic systems for forecasting, reasoning, and decision-making.

I am planning to apply for PhD intern positions in 2027 and welcome opportunities worldwide!
If you are interested in my work, please feel free to reach out: shifeng.xie@telecom‑paris.fr.


Research Experiences

Time Series Foundation Models

Huawei Paris Noah’s Ark Lab, France
February 2025 – August 2025

  • Designed classification time‑series foundation models and demonstrated that high‑performance pretraining is possible using only synthetic data.
  • First author of “CauKer: Classification Time Series Foundation Models Can Be Pretrained on Synthetic Data Only,” accepted by ICML 2025 Workshop on Foundation Models for Structured Data (Best Time Series Paper).
  • Collaborated with Ievgen Redko and researchers from Huawei Paris.

In‑Context Learning and Mixture of Experts

Stellantis, France
July 2024 – January 2025

  • Investigated in‑context learning and gradient descent in transformers and mixture‑of‑experts (MoE) models.
  • First author of “The Initialization Determines Whether In‑Context Learning Is Gradient Descent,” published in Transactions on Machine Learning Research (TMLR), 2025.
  • Worked closely with Rui Yuan, Simone Rossi and Thomas Hannagan.

Graph Neural Networks and Graph Representation Learning

Télécom Paris (Polytechnic Institute of Paris), France
December 2023 – September 2024

  • Conducted research on variational graph contrastive learning and subgraph Gaussian embedding for self‑supervised graph representation.
  • First author of “Variational Graph Contrastive Learning,” accepted by NeurIPS 2024 Workshop on Self‑Supervised Learning – Theory and Practice.
  • First author of “Subgraph Gaussian Embedding Contrast for Self‑Supervised Graph Representation Learning,” accepted by ECML‑PKDD 2025.
  • Supervised by Jhony H. Giraldo.

Image Processing and High‑Dynamic‑Range Reconstruction

Xidian University, China
February 2023 – September 2023

  • Developed FTUnet for single HDR image reconstruction.
  • First author of “FTUnet: Feature Transferred U‑Net for Single HDR Image Reconstruction,” accepted by ACM Multimedia Asia (MMA) 2023, oral presentation.
  • Advised by Liu Yi.

Data Twin and Intelligent Healthcare

Xidian University, China
April 2021 – September 2021

  • Researched the feasibility of intelligent healthcare based on digital twin and data mining.
  • First author of “Feasibility Study of Intelligent Healthcare Based on Digital Twin and Data Mining,” accepted by CISAI 2021.

Education

Master in Engineering – Signal Processing for Artificial Intelligence

Télécom Paris & Polytechnic Institute of Paris, France
August 2023 – Present

  • Current grade: 15.3 / 20.

Summer Exchange Program – Machine Learning

McGill University, Canada
July 2021 – August 2021

  • Achieved grade: A.

Bachelor of Engineering – Electronic Information Engineering

Xidian University, China
August 2019 – May 2023

  • GPA: 3.8 / 4.0.
  • Rank: 4 / 97.

Selected Projects

  • Pretraining tiny vision & language models (≈ 2 billion parameters): Trained mixture‑of‑experts models using 8 A100 80 GB GPUs on the C4 and ImageNet datasets.
  • Fine‑tuning MoE language models with permutation symmetries and LoRA (patented): Applied permutation symmetries and Low‑Rank Adaptation to MoE models (e.g., Mistral, DeepSeek, Qwen) to improve efficiency.
  • Unsupervised face recognition with PCA and ICA: Implemented PCA and ICA to extract and modify facial features, enhancing recognition accuracy.
  • Self‑supervised learning for medical image classification: Developed contrastive learning methods on the MedMNIST database to learn representations without annotations.
  • Neural network parameter diffusion (patented): Compressed experts from MoE models into a latent space via autoencoders and trained latent diffusion models to generate new experts.

Skills & Service

  • Languages: Chinese (native), English (C1), French (B2).
  • Programming: Python (PyTorch, TensorFlow, JAX, SciPy, Pandas), JavaScript, Java, C++ and C.
  • Hardware: Arduino, STM32, SolidWorks, VHDL, ARM and RISC.
  • Professional service: Reviewer for NeurIPS 2024 Workshop on Compression, COLM 2025, and NeurIPS 2025.