(CS 101) Projects in Machine Learning

Course Description

Prerequisite: CS 155 or equivalent

This is a project-based course for students looking to gain practical experience in machine learning. Students are expected to be proficient in basic machine learning. Students will work in groups. Each group will be provided a project topic to work on along with domain expert advisors. Alternatively, students can propose their own projects, subject to approval by course instructors.

Course Details

• Organizational meeting: 10am on Monday 10/1 in ANB 107.
• Organizational Meeting Slides
• SIGN UP SHEET
• 9-Unit lab course, no lectures or homeworks.
• Students will be assigned to a team and project at the beginning of the second week.
• Students will be required to provide weekly updates.
• Students are encouraged to consult instructors and teaching assistants for any technical issues that arise.

Instructor

Yisong Yue               yyue@caltech.edu
Omer Tamuz            omertamuz@gmail.com

Teaching Assistants

List of Projects

subject to change

Learning an Optimizer. Train a neural net to solve hard combinatorial optimization problems, such as Traveling Salesman.

• Mixed integer linear programs (MILPs) solved using local search heuristics such as branch-and-bound
• Treat local search heuristic as AI agent
• Partial solutions are states, and the goal is the find a state with high reward (feasible solution)
• Builds upon recent research @Caltech
• Relevant material: Learning to Search

Learning a Theorem Prover. Train a neural net to solve potentially undecidable problems, such as theorem proving.

• Treat theorem proving as a sequential decision making problem.
• Treat local search heuristic as AI agent
• Partial solutions are states, and the goal is the find a state with high reward (feasible solution)
• Builds upon recent research @Caltech by Lior Pachter

Earthquake Monitoring & Detection. Train a neural net to do better earthquaking detection.

• Caltech is responsible for all seismic monitoring in southern California
• Decades of data
• Recent success of deep learning PhaseLink
• Many exciting problems! Real-time estimation, de-noising to detect small-magnitude earthquakes, modeling evolution of earthquakes.

Bridge AI. Building a team of collaborating AI to learn to play bridge.

• Pair of cooperating AIs with partial information (can't see other players' hands)
• Similar to Poker, but requires cooperation.
• Requires reinforcement learning (can be learned along the way)
• Relevant material: CFR, Actor-Critic