学习⽬标
深度学习近年来已成为⼈⼯智能的核⼼技术之⼀，⼴泛应⽤于计算机视觉、⾃然语⾔处理、语⾳识别以及医学信号分析等领域。本次项⽬将带领学员深⼊探索卷积神经⽹络（CNN）、变换器（Transformers）等算法，并学习如何⾼效地处理时间序列信号，如⼼电图（ECG）数据。
本项⽬⾯向已具备深度学习基础的学员，故不包含基础课程讲解，直接进⼊项⽬实践。
核⼼⽬标是专注于对已有模型的改进与优化，通过在⾼质量私有数据集上进⾏测试验证，提升模型在实际场景中的表现。学习路径采⽤从⼩模型到⼤模型的渐进式⽅法：先对经典模型进⾏精读理解，再动⼿复现，最终在此基础上进⾏改进创新。
本项⽬不仅适合希望独⽴开展深度学习项⽬的学员，也适合希望掌握AI项⽬全流程的开发者。通过实际操作，学员将熟悉从数据预处理、模型设计、训练优化，到模型验证和应⽤部署的完整流程，并培养独⽴思考和解决问题的能⼒。项⽬中若取得精度提升等有效成果，团队可申请发表学术论⽂。
学习⽅式
项⽬实战
+ ⼩组讨论。本项⽬不安排常规理论课程，直接进⼊实战阶段。核⼼⽬标：
- AF检测精度提升：训练AI模型⾃动分析ECG信号，提⾼⼼房颤动（AFib）和⼼房扑动（AFL）的检测准确率，尤其是在数据质量较差（如噪声⼲扰）的场景下保持稳定表现
- 数据特性微调：针对不同数据特性（如⾼噪声数据）
设计专⻔的预处理或模型调优策略
- Duration-based检测：实现AF事件的起⽌时间定位，通过线段重叠⽐例（IoU）评估检测精度
项⽬资源：
⾼质量私有数据集：来⾃医疗公司的真实ECG数据，数据质量优于公开数据集
服务器算⼒⽀持：提供云端GPU服务器⽤于模型训练与实验
学习流程：
1）项⽬实战：学员将参与从零搭建深度学习模型的全过程，采⽤从⼩模型到⼤模型的渐进路径
——先精读经典模型架构，再动⼿复现，最后基于已有模型进⾏改进优化。
2）团队合作：学员将以⼩组形式合作完成项⽬（每组两⼈），模拟真实的AI研发流程，包括数据处理、建模、调优、验证等。
3）答疑点评：项⽬期间将邀请技术专家每两周进⾏⼀次答疑，
帮助解决疑难技术问题。
4）成果评估：项⽬结束后，每个⼩组将提交最终模型，并基于表现进⾏评分和反馈。
5）突破性成果：若项⽬中取得精度提升等有效成果，团队可申请发表 journal paper；否则撰写tech report。
阶段安排：
- 阶段1（第1-3周）：每周与医疗公司researcher召开2
次线上会议，理解项⽬背景、基本架构，获取公开数据集，进⾏模型精读与复现；
- 阶段2（第4-8周）：每周召开1-2次线上会议，开发改进模型并获取来⾃医疗公司的私有数据进⾏验证；
- 阶段3（第9-12周）：测试与验证阶段，每周召开1-2次线上会议，总结测试结果，根据反馈各组进⾏优化提升；
- 阶段4（第13-24周）：如果前12周有突破性成果，继续深入研究，并撰写学术论文争取发表。
学习时间
2026/03—2026/06，共12周，每周进⾏⼩组讨论及项⽬实践。
琪⽯会根据报名情况调整活动开始时间，请感兴趣的⼩伙伴提早报名，录满为止。
如果做出突出有效的成果，学习时间将延长至3到6个月，争取发学术论文。
学员组成
• 领队：⼀名（已到位）
• 组员：
6⼈（不含领队），不限专业，但需具备深度学习基础。由于项⽬规模有限，我们将以⾼标准的选拔流程邀请最适合参与的学员，共同营造⼀个⾼⽔平的学习交流环境。
报名⽅式
注册并登录陆琪⽯⽹站https://www.qishicpc.com, 在Courses中选择进⾏报名。
报名费100美元。报名者须参加⾯试评估，可能包括一至两轮⾯试，报名费⼀旦报名不予退还。
报名时请附上⾃⼰的简历和Cover Letter (简单说明为什么感兴趣、每周可投⼊时间等)。
报名链接：https://qishicpc.com/groupclass/6997d338083b23fa8a33baaa
报名二维码:
部分学习⼤纲:
Deep Learning Fundamentals, CNN Architecture, Time Series Data Processing, Transformer & Sequence Modeling, Data Preprocessing & Feature Extraction, Model Development (CNN, LSTM, Transformer), Training & Optimization, Model Evaluation & Deployment
具体学习资料暂略，录取后告知。
特别说明：Formal Course Fee $100是报名费。面试通过后才会收学费。
提早报名，录满为⽌。
2
nd Deep Learning Project Research Group
About
Deep learning has become a core technology in artificial intelligence, with wide-ranging applications in computer vision, natural language processing, speech recognition, and medical signal analysis. This project will guide participants in exploring key deep learning algorithms, including Convolutional Neural Networks (CNNs) and Transformers. A primary focus will be on efficiently processing time-series signals, such as electrocardiogram (ECG) data.
This project is designed for participants with prior deep learning experience and does not include introductory courses. Instead, it directly engages participants in hands-on projects. The core objective is to focus on improving and optimizing existing models, testing and validating them using high-quality private datasets to enhance their performance in real-world scenarios. The learning path adopts a progressive approach from small to large models: first, participants will thoroughly study and understand classic models, then reproduce them hands-on, and finally innovate and improve upon this foundation.
This project is suitable not only for those who wish to independently carry out deep learning projects, but also for developers seeking to master the entire AI project lifecycle. Through practical exercises, participants will gain familiarity with the complete process from data preprocessing, model design, training optimization, to model validation and deployment, while developing critical thinking and problem-solving skills. If effective outcomes such as improved accuracy are achieved during the project, the team may apply to publish academic papers.
Requirements
Application Fee: $100 (Applicants will undergo an evaluation process, which may involve two rounds of interviews. The fee is non-refundable.)
Resume and Cover Letter (including reasons for interest, availability per week, etc.) will be collected after you successfully applied.
Course Content
The program follows a hands-on project-based approach with group collaboration. There are no traditional lectures; participants will immediately begin working on real-world applications.
Primary objectives are:
AF Detection Accuracy Improvement: Train AI models to automatically analyze ECG signals, enhancing the detection accuracy of atrial fibrillation (AFib) and atrial flutter (AFL), especially maintaining stable performance even in situations with poor data quality (such as noisy interference).
1. Project Implementation: Participants will be involved in building a deep learning model from scratch, following a progressive approach from small to larger models.
—— Participants will first study classic models in depth, then replicate through practice, and finally improve and optimize based on existing models.
2. Team Collaboration: Participants will work in pairs, simulating an industry-like AI development workflow, including data collection, modeling, fine-tuning, and deployment.
3. Expert Q&A & Review: Guest experts will provide review sessions every two weeks during the project to address questions and provide feedback.
4. Performance Evaluation: Each team will submit a final model at the end of the project, with assessments and feedback based on performance.
5. Publication Opportunities: If a team achieves groundbreaking results, they will be encouraged to submit a journal paper. Otherwise, they will document findings in a technical report.
6. Project Phases:
- Phase 1 (Week 1–3): Weekly three online meetings with medical company researchers to understand the project background and basic architecture, acquire public datasets, and conduct in-depth study and replication of models.
- Phase 2 (Week 4–8): Weekly one to two online meetings to develop the model and gain access to private medical datasets.
- Phase 3 (Week 9–12): Testing & validation, with one to two online meetings every week to evaluate results and refine models based on feedback.
- Phase 4 (Week 13–24): If there are breakthrough results in the first 12 weeks, continue in-depth research and write academic papers for publication.
Project Resources:
High-quality private dataset: Real ECG data from medical companies, with quality superior to public datasets.
Server computing support: Cloud GPU servers provided for model training and experiments.
Member Composition
Team Lead: One (already assigned)
Team Members: Six (excluding the lead). Open to all fields of study, but deep learning proficiency is required. Due to project limitations, we will conduct a rigorous selection process to ensure a high-quality learning environment.
Course Schedule
2026/03—2026/06, totaling 12 weeks. Participants will engage in weekly group discussions and project work.
The start date will be adjusted based on enrollment. If interested, please apply early; enrollment will close when full.
If outstanding results are achieved, the study period may be extended to 3 to 6 months, with the goal of publishing academic papers.
Syllabus
Deep Learning Fundamentals, CNN Architecture, Time Series Data Processing, Transformer & Sequence Modeling, Data Preprocessing & Feature Extraction, Model Development (CNN, LSTM, Transformer), Training & Optimization, Model Evaluation & Deployment
More Specific study materials will be provided after admission.
Special Notice
The formal course fee of $100 is the registration fee. Tuition will only be collected after passing the interview.
Early application is recommended; enrollment will close once full.
琪石职业发展俱乐部
2014年成立于美国纽约，是一个为华人提供职业发展服务的平台，致力于成为金融和科技领域的“黄埔军校”。我们的高级会员就职于全球最顶尖的量化对冲基金和IT大厂。
会员专享
各类金融、科技领域的学习小组
线上线下networking活动
申请仅对会员开放的金融、IT行业内推职位
资深教练助力职业规划
琪石俱乐部已上线网站，请阅览我们网站来获取最新的“Activities”，“Materials”，“Positions”，“Hot topics” 以及其他更详细的信息网站地址是：https://www.qishicpc.com/
俱乐部目前开放各地对外微信大群，欢迎注册俱乐部网站，熟悉和联系我们，也欢迎联系我们加入微信大群
如有疑问，也欢迎发电子邮件至qishicpc@gmail.com
各位琪石的小伙伴，一起在文末留言你的想法/问题吧！
也欢迎分享给可能感兴趣的朋友们。
未来我们一起同行，在不断地学习和尝试中，成就更好的自己。
