CS 341: Machine Learning

How does Netflix learn what movies a person likes? How do computers read handwritten addresses on packages, or detect faces in images? Machine learning is the practice of programming computers to learn and improve through experience, and it is becoming pervasive in technology and science. This course will cover the mathematical underpinnings, algorithms, and practices that enable a computer to learn. Topics will include supervised learning, unsupervised learning, evaluation methodology, and Bayesian probabilistic modeling. Students will learn to program in MATLAB and apply course skills to solve real-world prediction and pattern recognition problems. Programming intensive. Instructor: Dan Sheldon  dsheldon (at) mtholyoke (dot) edu Meeting time: Mondays and Wednesdays from 2:40–3:55 Location: Kendade, Room 303 ella: https://ella.mtholyoke.edu/portal/site/COMSC-341-01-FA12 Contacting me: Office hours: Monday 4-5 p.m. / Tuesday 7:15-8:30 p.m. / by appointment, Clapp 200 Email: Email is a very good way to get hold of me for administrative items. I will typically respond within 24 hours (48 hours on weekends), if not much sooner, to emails about administrative items. For help on assignments, please come to office hours. Note: Because I have a joint appointment, I will be at UMass on non-class days, which is why office hours are held on the same days as class. If this situation is not working for you, let me know and we'll work something out! Prerequisites: The two official prerequisites are Computer Science 211 Data Structures and Math 232 Discrete. Together these should ensure what really matters: (1) you should be comfortable programming in some language, and (2) you are not afraid of math. A word about math: machine learning is inherently mathematical. This course will utilize some material from basic calculus, probability and statistics, and linear algebra. I will teach or review the necessary material, so you do not need to know it in advance. However, you should be comfortable with math.

luckey_sun | Flickr via Creative Commons

Required Textbook and Other Resources

The required textbook is Artificial Intelligence: A Modern Approach, Third Edition, by Russell and Norvig. It is important to get the third edition (cover shown at right), which has substantial changes to the machine learning content of the book. The book is not available at Odyssey Bookstore. You can get it from Amazon ($111.99). Various digital versions are also available (e.g. from CourseSmart for $63.99). This course will draw primarily from Chapters 13–15, 18, and 20, which discuss probabilistic reasoning and learning. The rest of the book is an outstanding and widely used text on Artificial Intelligence (AI); if you wish to study AI in the future, consider purchasing a hard copy. A number of other resources are available: Andrew Ng teaches an outstanding (and free) online machine learning course at Coursera. As of this writing, you can preview any of the lectures. Parts of this course will be based on those lectures—I encourage you to check them out. Trevor Hastie, Rob Tibshirani, and Jerry Friedman, Elements of Statistical Learning, Second Edition, Springer, 2009. This is an authoritative (but advanced) book on statistical machine learning. It is freely downloadable from the website. Alpaydin, Introduction to Machine Learning, Second Edition, MIT Press, 2010. An accessible introductory text to machine learning. Bishop, Pattern Recognition and Machine Learning, Springer, 2006. The Bishop and Alpaydin books, as well as one copy of Russell and Norvig, are on reserve at the MHC library.

Course Purpose and Goals

• Mathematical tools: probability, matrix and vector manipulation, basic optimization
• Machine learning principles: problem formulations, notation, overfitting, regularization, quantifying uncertainty
• Methodology: evaluating machine learning algorithms, statistical concepts, diagnosing and controlling overfitting
• Applications: discussion and hands-on experience applying machine learning algorithms to data, including some of the "messy stuff" such as data preparation and feature engineering
• Fun: teaching computers to do amazing things!

Tentative List of Topics

• Intro: What is Machine Learning?
• Supervised Learning
  • Linear regression in one variable
  • Linear algebra review
  • Intro to MATLAB
  • Multivariate linear regression
  • Logistic regression
  • Overfitting, regularization, model complexity
  • Methods: feature design
  • Decision Trees
  • Neural Nets
  • Methods: experiment design
• Unsupervised Learning
  • Clustering (hierarchical; k-means)
  • Dimensionality reduction (PCA; Eigenfaces)
• Probabilistic Methods
  • Probability review
  • Bayesian networks
  • Hidden Markov models
  • Statistical learning
  • Naive Bayes
  • Learning hidden Markov models
  • Learning with incomplete data: the EM algorithm

As the term unfolds, I will update the actual course schedule with the topics covered as well as readings, assignments, and other resources.

Course Policies

Coursework

• Homework assignments will be a mix of written exercises, programming, and experimentation. These will be issued and due on Wednesdays. I encourage you to collaborate on these assignments. There is no limit to the number of students that may work together, however every student must write their own code, run their own experiments, and write their own solutions. Also, I highly encourage each student to first attempt the problems on their own, especially for the shorter exercises that are designed to test and reinforce concepts taught in class. Please write the names of all collaborators at the beginning of the written portion of the submission. Failure to follow any of these guidelines is a violation of the honor code.
  
  Submission: digital files (code, etc.) should be submitted to ella before class begins on the due date. Written work can be turned in either by ella as pdf files, or in hard copy at the beginning of class.
  

• The course project is your chance to explore. The ideal project will creatively apply machine learning or data mining skills learned in this class to a domain that interests you: for example, you may wish to solicit a scientific data set from another Mt. Holyoke faculty member and try out some machine learning methods on it; or, you may find a dataset related to an hobby of passion of yours, and explore it using unsupervised machine learning methods; or, you may decide to try out a machine learning algorithm that is not covered in this class. More details about the project, including suggested topics, datasets, and due dates, will appear later.

  Students may work in groups of 2–3 for the project. I encourage you to select your own groups, but I can help make matches as well. Every student in a project group will receive the same grade (barring extreme circumstances, in which case, please come to me to work out a different solution). Submitted work for the project will consist of: a project proposal, weekly status reports, and a final report written in the format of a short scientific paper.
  

Late submissions

Each student will have five free late days to be used on homeworks and project milestones (charged to every member of the project group). Each late day buys exactly 24 hours from the original due date at the beginning of class (so 24.5 hours = 2 late days). If you use up your late days, you will be penalized 33% of the assignment's value for each day or fraction thereof that it is late (0–24 hours = 33% penalty; 24–48 hours = 66% penalty; 48+ hours = no credit). Since I will not generally be at Mt. Holyoke on non-class days, late work should be submitted via ella. It is your responsibility to scan any written work and submit it in pdf form.

Late days may not be used on the final project report.

• Homework: 60%
• Project: 30%
• Class participation: 10%