University of Bern
M.Sc. in Computer Science.
Currently in the second year of the program. During my studies, I have been collaborating on research projects with the Computer Vision Group, focusing on optimization in deep learning.
Supervised by Prof. Dr. Paolo Favaro.

Xi’an Jiaotong University
B.Eng. in Software Engineering, advised by Prof. Jihua Zhu.
Thesis: Music Generation Based on Large Language Models.

K-Score: Kalman Filter as a Principled Alternative to Reward Normalization in Reinforcement Learning
Proposed a lightweight Kalman-filter-based reward normalization framework to replace heuristic Z-score normalization in policy gradient reinforcement learning. Developed both fixed and adaptive 1D Kalman filter modules for online return estimation, enabling dynamic adaptation to non-stationary reward distributions. Demonstrated improved convergence speed and stability on CartPole and LunarLander using REINFORCE, Actor-Critic, GAE, and PPO. Worked jointly with Quanxi Li; submitted and presented at the New in ML 2025 Workshop.

Music Generation with Large Language Models
Conducted undergraduate thesis research on symbolic music generation using deep learning, focusing on recurrent networks, Transformers, and LLMs. Evaluated models on the MAESTRO dataset using both MIDI and waveform outputs; visualized training and validation performance. Released code on GitHub: DL4Music.

Multilingual Document Layout Analysis on Public Procurement Data
Conducted few-shot layout analysis on multilingual CFT documents (German, French, Italian) provided by IntelliProcure. Benchmarked YOLOv10 and Transformer-based (LayoutXLM, LiLT) models under limited labeled data. Built pipeline for PDF preprocessing, region extraction, comparison, and qualitative evaluation. Evaluated model accuracy and inference time, with oral presentation of results.

Time Series Forecasting on Real Bank Transaction Data
Collaborated with Prof. Lufei Huang on a banking project using proprietary transaction data. Built a CNN-based model to forecast 7×7 matrix-style flows; achieved cosine similarity > 0.73. Developed full pipeline with preprocessing, normalization, visualization, and inference logging.

Deep Learning Notes Organization and Rewriting
Reorganized and rewrote materials from multiple deep learning courses into cohesive documents. Completed 15 assignments from Hung-Yi Lee’s curriculum with structured notes. Published on GitHub, gaining 100+ stars: Repo Link. In Spring 2025, began English rewriting project based on Prof. Paolo Favaro (University of Bern)'s deep learning course, integrating key ideas into LaTeX-based materials.

KOALA++: Efficient Kalman-Based Optimization of Neural Networks with Gradient-Covariance Products An extension of KOALA, a neural network optimization algorithm based on Kalman filtering, with implicit full weights covariance matrix.