(CS 101) Projects in Machine Learning

Instructor

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

Teaching Assistants

List of Projects

subject to change

Visualizing & Analyzing twitter Data. Mentored by Mike Alvarez. Analyzing twitter data is extremely challenging, given the streaming nature of the data as well as the sheer volume. This project will be involve creating a pipeline to visualize analysis of twitter data in an easy to view way. Examples include simple analyses such as creating distributions of twitter words by geography, to more complicated ones such as visualzing latent-factor modeling analysis of twitter data.

Data Forensics for Voter Registration Data. Mentored by Mike Alvarez. Develop tools to collect and analyze voter registration data. A number of states (e.g., NC and OH) post their entire state’s voter registration and voter history file online, with weekly updates. One can use machine learning to identify simple problems such as duplicate entries and minor errors, or challenging errors such as ballot stuffing. See this paper for an early example.

Computational humor. Write a program that makes people laugh (with it, not at it). Some options include: an automatic meme generator, a twitter bot, and a reaction gif suggester.

Adaptive Teaching. Teach children the alphabet / basic reading / basic math by learning on the fly what they know and don't know and choosing learning tasks accordingly.

Minimum Intelligent Signal Test. This test is a variant of the Turing test, in which the a tester asks a responder questions, and the responder is only allowed to answer "yes" or "no". Build a system in which users are randomly assigned to be either testers or responders, and testers chat with responders who are randomly chosen to be either machines or humans. Learn from the answers given by human responders to train machine responders.

AI for Quantum Chess (Requires CS 159). Mentored by Spiros Michalakis and Chris Cantwell. Quantum Chess is a new game developed by Caltech, USC, and industry partners. Students must be familiar with concepts such as Monte Carlo tree search.

The goal of the project is to develop an AI that can play the game of Quantum Chess against itself and other human players. Quantum Chess is a variant of chess that allows the player to move a piece in superposition. After the move you can consider there to be two possible boards, one in which the move was successful and one in which the piece didn't move at all. The player is presented with a view of the board that shows positions of pieces along with the associated probability that a measurement will find it in that position. The quantum moves obey the mathematics of quantum mechanics, so the player can experience phenomena such as entanglement and interference during gameplay. There is also a rule for when a measurement occurs, thus determining which of the possible board states are “real”, and in the process changing the state of the entire board in ways that can give a massive advantage to a player that was expected to lose.

There are a number of challenges to building a Quantum Chess A.I. The existence of superposition alone, means that the branching factor for Quantum Chess is much higher than that of Chess. Moreover, the fact that the player is only presented with probability information means that predicting interference effects can be difficult. Finally, the state of the board after a measurement is impossible to determine a priori. This implies that a simple tree search algorithm will need to be modified in order to work with Quantum Chess. For example, the A.I. may need to take into account information about the expected value of every piece on the board (that is, if a Queen is worth 6 points in regular chess, but is split over two different locations, then the value of each queen is expected to be 0.5*6=3), when evaluating the different board configurations. At a more advanced level, the AI may need to have access to the whole history state of the game, if it plans to make use of interference and entanglement between pieces, in order to develop more complex strategies.

We expect this to be a very rewarding project, playing an important role in the success of Quantum Chess as a game and as an educational tool. Moreover, the team working on this project will be the first team ever to develop a Quantum A.I., a very exciting frontier for exploration within the field of quantum games and quantum computing.

AI for Classical Games (Requires CS 159). The project will usef interactive machine learning to train AI systems to play various classical games. Students must be familiar with concepts such as Monte Carlo tree search. Candidate games include:

• Atari
• Board Games

(No Longer Available) Machine learning techniques applied to data from Mars Science Laboratory (MSL). Mentored by Michela Munoz Fernandez.

• Project 1: Development of an anomaly detector system for the Mars Science Laboratory (MSL): Generation of models trained on historical data that will learn to recognize both nominal and off-nominal behaviors, including never-before-seen anomalies.
• Project 2: Wheel crack measurements of the Mars Science Laboratory (MSL): The machine learning system would have to recognize what a crack looks like in the images provided and then measure it to track the degradation of the wheels over time.

Machine Learning for Sports. Mentored by Hoang Le. Develop ML interface with the RoboCup simulator. Work on ML algorithms to learn defensive formations in soccer. May involve tracking data from professional games.

Machine Learning for Mobile Platforms. Mentored by Grant van Horn. Mobile devices offer unique challenges to machine learning pipelines: not only do the systems still need to achieve low error, but they also must satisfy strict memory and performance constraints. Many state of the art deep convolutional neural networks used for image classification do not readily fit on mobile devices. The goal of this project is to develop a computer vision model for bird species classification that can run efficiently on mobile phones that range from 512 MB to 2+ GB of memory while gracefully degrading in performance. Possible directions:

• Compression techniques for neural nets
• Experiment with reducing layer size
• Quantizing networks

Improve Open Source ML Platforms. Develop extensions of existing ML platforms (such as deep learning). Examples include:

• Improving parallel machine learning.