Lin Li, Ph.D.

Data Scientist & Machine Learning Engineer

I am a Data Scientist with a Ph.D. in Applied Mathematics and deep expertise in applied statistics, computer vision, and deep learning. I'm passionate about building scalable machine learning systems that solve complex real-world problems. My work spans from algorithm design to deployment, with a strong interest in medical imaging, 3D reconstruction, and high-performance data pipelines.

LinkedIn Google Scholar

Professional Experience

Building innovative solutions across medical imaging and autonomous driving

Alphatec Spine(2023-present)

I led the redesign of the C-arm calibration process, reducing installation time by over 75% and increasing system pass rates by more than 60%. I also spearheaded the development of an intra-operative image alignment (IOA) system, playing a key role in its FDA 510(k) submission and alpha launch. My contributions have enabled faster, more reliable surgical workflows and improved clinical outcomes.

Learn More

TuSimple(2020-2023)

I built and deployed deep learning models for 3D perception in autonomous vehicles. I developed large-scale data pipelines, monocular and stereo 3D estimation models, and real-time APIs, significantly improving depth estimation accuracy and automation in challenging environments.

Learn More

Presentations

Technical talks and research reviews

EMMA Paper reading

EMMA: End-to-end multimodal model built on Gemini for autonomous driving
State-of-the-art planning on nuScenes, competitive on WOMD
Competitive 3D detection on camera-primary WOD benchmark
Generalist model that improves across multiple tasks via co-training

DeepSeek Introduction

Comprehensive overview of DeepSeek-R1 architecture and training innovations
Key technical breakthroughs: Mixture-of-Experts (MoE), Multi-head Latent Attention (MLA), and Multi-Token Prediction (MTP)
Achieved performance comparable to OpenAI o1 with only 1/20th of the computational resources
Advanced training pipeline utilizing Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO)
Explored practical applications including automated regulatory submissions, financial analysis, and report generation

EPro-PnP

Proposed a probabilistic PnP layer for general end-to-end pose estimation via learnable 2D-3D correspondences
EPro-PnP can easily reach top-tier performance for 6DoF pose estimation by simply inserting it into the CDPN framework

Proximal Policy Optimization(PPO)

Background: Fundamental concepts of Reinforcement Learning
Proximal Policy Optimization (PPO): Core algorithm
Group Relative Policy Optimization (GRPO): Modern variant
Group Sequence Policy Optimization (GSPO): Sequence-level optimization