Class Project Information

Teams

The final project will be done in teams of (up to) 2 people. Please fill in the team assignment form before Nov 30.

Important Milestones

Project Topic

The project topic can be very open-ended. You may choose any topic you are interested in (such as your research area) as long as it pertains to the course material. You are encouraged to explore advanced machine learning techniques in supervised, unsupervised and reinforcement learning.

Usually, your project will fall into one of the following categories:

• Application: apply a machine learning method to solve a specific problem, e.g.

  • Machine learning for AI applications in computer vision, NLP, speech recognition, etc

  • Use a learning-based approach for physical systems, such as energy, material and optics

• Algorithm: develop a machine learning method that has better performance, e.g.

  • Improve the robustness of an algorithm under noisy data

  • Make a machine learning method more efficient

• Theory: theoretical problems, e.g.

  • Theoretical analysis of a specific learning algorithm regarding its generalization, consistency, or stability properties

  • In-depth tutorial of a particular theoretical topic

A good project will showcase your ability to clearly define a problem, understand the motivation and technical details of your method, and complete significant implementation with real data.

Grading

The final project will count for 25% of your final grade. It will be graded on a scale of 100 points, which include:

• Proposal: 15 points

• Discussion with class staff: 5 points

• Final Poster (Instructor and Invited Guests): 20 points

• Final Poster (Other teams): 10 points

• Final Report Grades: 50 points

Sample Projects from Previous Years

• Deep Hypergraph Convolutional Network for Paper Classification, Yifei Zhu and Guanzi Chen (algorithm & theory)

• Machine Learning in Casual Inference: A Survey and Application in Economics, Pengxiang Zhou and Yanyu Chen (application)

• An Application of Machine Learning on Predicting CO2 Reduction Electrocatalysts, Chen Liang and Yuhang Zhang (application)

• Live Gradient Compensation for Evading Stragglers in Distributed Learning Jian Xu (algorithm)

• Debugging Neural Networks, Riccardo Mattesini, Sebastian Beetschen and Bunchalit Eua-arporn (algorithm)

Common Pitfalls

• Simply run a model from an existing work on a slightly different dataset: think of your contribution in at least one of the three catogries discussed in the Topics section

• Topic is too broad or too ambitious: reduce the project scope

• Project relies heavily on data availability/quality: use available datasets