Yanli Liu Email and Phone Number

2024 Half-year Outstanding Employee AwardCore contributor to establishing Shein's AI-driven product review system. Daily business value: ~32k USD; Daily potential lawsuit cost avoided: ~2.9m USD. The values are rapidly increasing.The system prevents infringements (e.g., brand names, logos) and taboos (e.g., religious and political symbols) on Shein products. I am the core contributor to problem definition, system design, model training, system deployment, and system maintenance. In terms of tech stack, I worked on:1. Vision language models. Retrieval Augmented Generation (RAG). Supervised finetuning (SFT), Parameter-efficient finetuning (PEFT). Model quantization. Tensor Parallel. Pipeline Parallel.2. Image retrieval. Extreme long-tail distributions. High-quality training data construction. Efficient multiprocessing.3. Scene text recognition. Curved texts. Texts in artistic styles.I also communicate with the stakeholders and continuously improve the user experience.

2022 Excellent Employee Award (4 out of 120 employees).Currently, I am leading projects on 1) Guided diffusion models for controlled image generation and editing, the goal is to generate novel viewpoints while maintaining editability by text prompts, without using NeRF representations. 2) Highly accurate and lightweight image segmentation under challenging lighting conditions and human poses.Previously, I led the following research&productization projects:1. Image enhancement for smartphone cameras with light computation cost (30ms to process a 12-megapixel photo), deployed on Realme flagship smartphones in 2023 (e.g., Realme GT Neo 5, Realme 11 series, etc).2. Model compression methods, including neural network pruning and quantization for deployment on hardware platforms. Significant performance boost for OPPO's under-screen camera.3. Image translation and neural style transfer models, a lightweight GAN that takes ~250ms to process a 12-megapixel photo on smartphones.4. Lightweight add-on module to significantly boost the performance of CNNs and vision transformers with at most 10% more FLOPs and runtime.

INFORMS Optimization Society Best Student Paper Prize Honorable Mention (top 6)UCLA Dissertation Year Fellowship (top 15%)Advisor: Prof. Wotao YinResearch Interests:Machine Learning: provable algorithms for deep learning, optimization landscape of reinforcement learning, efficient algorithms for federated learning.Optimization: effective preconditioning for first-order methods, saddle point avoidance, detection and prevention of pathologies.

A paper on the role of momentum in SGD published at NeurIPS 2020, >100 citations in 2.5 years.Detailed achievements: 1. New convergence theory for (multistage) SGD with momentum2. A clarification of the acceleration from multistage parameter tuning3. A comparison of Adam with Gauss-Newton methods, which suggests better preconditioners4. A paper published at NeurIPS 2020

Best presentation award (6 out of 89)UCLA Cross-disciplinary Scholars in Science and Technology (CSST) is one of the best research-oriented summer programs(REU) for China mainland undergraduates.Conduct research in applied mathematics and prepare for graduate application.