CS 189/289A: Introduction to Machine Learning (original) (raw)

	CS 189/289A Introduction to Machine Learning Jonathan Shewchuk (Please send email only if you don't want the TAs to see it; otherwise, usePiazza.)Spring 2017Mondays and Wednesdays, 6:30–8:00 pm2050 Valley Life Sciences BuildingMy office hours:Mondays, 5:10–6 pm, 529 Soda Hall,Wednesdays, 9:10–10 pm, 411 Soda Hall, and by appointment.(I'm usually free after the lectures too.)

This class introduces algorithms for learning, which constitute an important part of artificial intelligence.

Topics include

• classification: perceptrons, support vector machines (SVMs), Gaussian discriminant analysis (including linear discriminant analysis, LDA, and quadratic discriminant analysis, QDA), logistic regression, decision trees, neural networks, convolutional neural networks, nearest neighbor search;
• regression: least-squares linear regression, logistic regression, polynomial regression, ridge regression, Lasso;
• density estimation: maximum likelihood estimation (MLE);
• dimensionality reduction: principal components analysis (PCA), random projection, latent factor analysis; and
• clustering: _k_-means clustering, hierarchical clustering, spectral graph clustering.

Useful Links

• See the schedule of class and discussion section times and rooms.
• Schedule of office hours.
• Access the CS 189/289APiazza discussion group.
• If you want an instructional account, you canget one online. No more paper forms. Go to the same link if you forget your password or account name.
• Photos of all your teaching assistants.

Prerequisites

• Math 53 (or another vector calculus course),
• Math 54 or 110 (or another linear algebra course),
• CS 70 (or another probability course). You should take these prerequisites quite seriously: if you don't have them, I strongly recommend not taking CS 189.

If you want to brush up on prerequisite material, Stanford's machine learning class provides nice reviews oflinear algebraandprobability theory. Here's ashorter summary of math for machine learning written by our TA Garrett Thomas. Other suggestions for review material appear inthis Piazza post.

Textbooks

Both textbooks for this class are available free online. Hardcover and eTextbook versions are also available.

• Gareth James,Daniela Witten,Trevor Hastie, andRobert Tibshirani,An Introduction to Statistical Learning with Applications in R, Springer, New York, 2013. ISBN # 978-1-4614-7137-0.See Amazon for hardcover or eTextbook.
• Trevor Hastie,Robert Tibshirani, andJerome Friedman, The Elements of Statistical Learning: Data Mining, Inference, and Prediction, second edition, Springer, 2008.See Amazon for hardcover or eTextbook.

Homework and Exams

You have a total of 5 slip days that you can apply to your semester's homework. We will simply not award points for any late homework you submit that would bring your total slip days over five.

We have extended the Homework 1 due date by three days.Homework 1is due Monday, January 30 at 11:59 PM.

Homework 2is due Monday, February 13 at 11:59 PM.

Homework 3is due Monday, February 27 at 11:59 PM.

Homework 4is due Monday, March 13 at 11:59 PM.

Homework 5is due Monday, March 27 at 11:59 PM.

Homework 6is due Friday, April 14 at 11:59 PM.

Homework 7is due Friday, April 28 at 11:59 PM.

The CS 289A Projecthas a proposal due Monday, April 10. The video and final report are due Friday, May 5.

The Midterm takes place on Monday, March 13, at either 6:30–8:00 PM in 2050 VLSB or 7:00–8:30 PM in Pimental Hall, depending on which room we assign you to. Previous midterms are available:Spring 2013,Spring 2014,Spring 2015,Fall 2015,Spring 2016.

The Final Exam takes place on **Monday, May 8, 3–6 PM,**in the RSF Fieldhouse. CS 189 is in exam group 3. Previous final exams are available. Without solutions:Spring 2013,Spring 2014,Spring 2015,Fall 2015,Spring 2016,Spring 2017. With solutions:Spring 2013,Spring 2014,Spring 2015,Fall 2015,Spring 2016,Spring 2017.

Lectures

Now available:The complete semester's lecture notes (with table of contents and introduction).

Lecture 1 (January 18): Introduction. Classification, training, and testing. Validation and overfitting. Read ESL, Chapter 1. My lecture notes (PDF). The screencast.

Lecture 2 (January 23): Linear classifiers. Predictor functions and decision boundaries. The centroid method. Perceptrons. Read parts of the WikipediaPerceptron page. Optional: Read ESL, Section 4.5–4.5.1. My lecture notes (PDF). The screencast.

Lecture 3 (January 25): Gradient descent, stochastic gradient descent, and the perceptron learning algorithm. Feature space versus weight space. The maximum margin classifier, aka hard-margin support vector machine (SVM). Read ISL, Section 9–9.1. My lecture notes (PDF). The screencast.

Lecture 4 (January 30): The support vector classifier, aka soft-margin support vector machine (SVM). Features and nonlinear decision boundaries. Read ESL, Section 12.2 up to and including the first paragraph of 12.2.1. My lecture notes (PDF). The screencast.

Lecture 5 (February 1): Machine learning abstractions: application/data, model, optimization problem, optimization algorithm. Common types of optimization problems: unconstrained, constrained (with equality constraints), linear programs, quadratic programs, convex programs. Optional: Read (selectively) the Wikipedia page onmathematical optimization. My lecture notes (PDF). The screencast.

Lecture 6 (February 6): Decision theory: the Bayes decision rule and optimal risk. Generative and discriminative models. Read ISL, Section 4.4.1. My lecture notes (PDF). The screencast.

Lecture 7 (February 8): Gaussian discriminant analysis, including quadratic discriminant analysis (QDA) and linear discriminant analysis (LDA). Maximum likelihood estimation (MLE) of the parameters of a statistical model. Fitting an isotropic Gaussian distribution to sample points. Read ISL, Section 4.4. Optional: Read (selectively) the Wikipedia page onmaximum likelihood. My lecture notes (PDF). The screencast.

Lecture 8 (February 13): Eigenvectors, eigenvalues, and the eigendecomposition. The Spectral Theorem for symmetric real matrices. The quadratic form and ellipsoidal isosurfaces as an intuitive way of understanding symmetric matrices. Application to anisotropic normal distributions (aka Gaussians). Read Chuong Do's notes on the multivariate Gaussian distribution. My lecture notes (PDF). The screencast.

Lecture 9 (February 15): Anisotropic normal distributions (aka Gaussians). MLE, QDA, and LDA revisited for anisotropic Gaussians. Read ISL, Sections 4.4 and 4.5. My lecture notes (PDF). The screencast.

February 20 is Presidents' Day.

Lecture 10 (February 22): Regression: fitting curves to data. The 3-choice menu of regression function + loss function + cost function. Least-squares linear regression as quadratic minimization and as orthogonal projection onto the column space. The design matrix, the normal equations, the pseudoinverse, and the hat matrix (projection matrix). Logistic regression; how to compute it with gradient descent or stochastic gradient descent. Read ISL, Sections 4–4.3. My lecture notes (PDF). The screencast.

Lecture 11 (February 27): Newton's method and its application to logistic regression. LDA vs. logistic regression: advantages and disadvantages. ROC curves. Weighted least-squares regression. Least-squares polynomial regression. Read ISL, Sections 4.4.3, 7.1, 9.3.3; ESL, Section 4.4.1. Optional: here isa fine short discussion of ROC curves—but skip the incoherent question at the top and jump straight to the answer. My lecture notes (PDF). The screencast.

Lecture 12 (March 1): Statistical justifications for regression. The empirical distribution and empirical risk. How the principle of maximum likelihood motivates the cost functions for least-squares linear regression and logistic regression. The bias-variance decomposition; its relationship to underfitting and overfitting; its application to least-squares linear regression. Read ESL, Sections 2.5 and 2.9. Optional: Read the Wikipedia page onthe bias-variance trade-off. My lecture notes (PDF). The screencast.

Lecture 13 (March 6): Ridge regression: penalized least-squares regression for reduced overfitting. How the principle of maximum a posteriori (MAP) motivates the penalty term (aka Tikhonov regularization). Subset selection. Lasso: penalized least-squares regression for reduced overfitting and subset selection. Read ISL, Sections 6–6.1.2, the last part of 6.1.3 on validation, and 6.2–6.2.1; and ESL, Sections 3.4–3.4.3. Optional: This CrossValidated page onridge regression is pretty interesting. My lecture notes (PDF). The screencast.

Lecture 14 (March 8): Kernels. Kernel ridge regression. The polynomial kernel. Kernel perceptrons. Kernel logistic regression. The Gaussian kernel. Optional: Read ISL, Section 9.3.2 and ESL, Sections 12.3–12.3.1 if you're curious about kernel SVM. My lecture notes (PDF). The screencast.

The Midtermtakes place on Monday, March 13, at either 6:30–8:00 PM in 2050 VLSB or 7:00–8:30 PM in Pimental Hall, depending on which room we assign you to. You are permitted one “cheat sheet” of letter-sized (8½" × 11") paper.

Lecture 15 (March 15): Decision trees; algorithms for building them. Entropy and information gain. Read ISL, Sections 8–8.1. My lecture notes (PDF). The screencast.

Lecture 16 (March 20): More decision trees: multivariate splits; decision tree regression; stopping early; pruning. Ensemble learning: bagging (bootstrap aggregating), random forests. Read ISL, Section 8.2. My lecture notes (PDF). The screencast.

Lecture 17 (March 22): Neural networks. Gradient descent and the backpropagation algorithm. Read ESL, Sections 11.3–11.4. Optional: I've heard positive recommendations for Welch Labs' video tutorialNeural Networks Demystified on YouTube. Also of special interest is this Javascriptneural net demothat runs in your browser. My lecture notes (PDF). The screencast.

March 27–31 is Spring Recess.

Lecture 18 (April 3): Neuron biology: axons, dendrites, synapses, action potentials. Differences between traditional computational models and neuronal computational models. Backpropagation with softmax outputs and logistic loss. Unit saturation, aka the vanishing gradient problem, and ways to mitigate it. Heuristics for avoiding bad local minima. Optional: Try out some of the Javascript demos onthis excellent web page—and if time permits, read the text too. The first four demos illustrate the neuron saturation problem and its fix with the logistic loss (cross-entropy) functions. The fifth demo gives you sliders so you can understand how softmax works. My lecture notes (PDF). The screencast.

Lecture 19 (April 5): Heuristics for avoiding bad local minima. Heuristics for faster training. Heuristics to avoid overfitting. Convolutional neural networks. Neurology of retinal ganglion cells in the eye and simple and complex cells in the V1 visual cortex. Read ESL, Sections 11.5 and 11.7. Optional: A fine paper on heuristics for better neural network learning isYann LeCun, Leon Bottou, Genevieve B. Orr, and Klaus-Robert Müller, “Efficient BackProp,” in G. Orr and K.-R. Müller (Eds.),Neural Networks: Tricks of the Trade, Springer, 1998. Also of special interest is this Javascriptconvolutional neural net demo that runs in your browser. My lecture notes (PDF) andsome slides about the V1 visual cortex and ConvNets(PDF). The screencast.

Lecture 20 (April 10): Unsupervised learning. Principal components analysis (PCA). Derivations from maximum likelihood estimation, maximizing the variance, and minimizing the sum of squared projection errors. Eigenfaces for face recognition. Read ISL, Sections 10–10.2 and the Wikipedia page onEigenface. My lecture notes (PDF). The screencast.

Lecture 21 (April 12): The singular value decomposition (SVD) and its application to PCA. Clustering: _k_-means clustering aka Lloyd's algorithm;_k_-medoids clustering; hierarchical clustering; greedy agglomerative clustering. Dendrograms. Read ISL, Section 10.3. My lecture notes (PDF). The screencast.

Lecture 22 (April 17): Spectral graph partitioning and graph clustering. Relaxing a discrete optimization problem to a continuous one. The Fiedler vector, the sweep cut, and Cheeger's inequality. The vibration analogy. Greedy divisive clustering. The normalized cut and image segmentation. Read my survey of Spectral and Isoperimetric Graph Partitioning, Sections 1.2–1.4, 2.1, 2.2, 2.4, 2.5, and optionally A and E.2. For reference: Jianbo Shi and Jitendra Malik,Normalized Cuts and Image Segmentation, IEEE Transactions on Pattern Analysis and Machine Intelligence22(8):888–905, 2000. My lecture notes (PDF). The screencast.

Lecture 23 (April 19): Learning theory, range spaces (aka set systems), dichotomies, the shatter function, and the Vapnik–Chervonenkis dimension. Read Andrew Ng'sCS 229 lecture notes on learning theory. My lecture notes (PDF). The screencast.

Lecture 24 (April 24): Graph clustering with multiple eigenvectors. Latent factor analysis (aka latent semantic indexing). Nearest neighbor classification and its relationship to the Bayes risk. The geometry of high-dimensional spaces. The exhaustive algorithm for _k_-nearest neighbor queries. Random projection. Read ISL, Section 2.2.3. Optional: Read the Wikipedia page onlatent semantic analysis. For reference: Andrew Y. Ng, Michael I. Jordan, Yair Weiss, On Spectral Clustering: Analysis and an Algorithm, Advances in Neural Information Processing Systems 14 (Thomas G. Dietterich, Suzanna Becker, and Zoubin Ghahramani, editors), pages 849–856, the MIT Press, September 2002. My lecture notes (PDF). The screencast.

Lecture 25 (April 26): Speeding up nearest neighbor queries. Voronoi diagrams and point location.k_-d trees. Application of nearest neighbor search to the problem of_geolocalization: given a query photograph, determine where in the world it was taken. For reference: the best paper I know about how to implement a _k_-d tree is Sunil Arya and David M. Mount,Algorithms for Fast Vector Quantization, Data Compression Conference, pages 381–390, March 1993. For reference: the IM2GPS web page, which includes a link to the paper. My lecture notes (PDF). The screencast.

Special Lecture by Daylen Yang and Raul Puri (May 4): Deep learning. Recurrent neural networks. Image captioning and visual question answering. (This lecture is not testable on the Final Exam.) Their talk slides (PDF).

The Final Exam takes place onMonday, May 8, 3–6 PM in the RSF Fieldhouse. (CS 189 is in exam group 3.)

Discussion Sections and Teaching Assistants

Here are photos of all your teaching assistants!

Sections begin to meet on January 24. A schedule will appear here.

Grading

• 40% for homeworks.
• 20% for the Midterm.
• CS 189: 40% for the Final Exam.
• CS 289A: 20% for the Final Exam.
• CS 289A: 20% for a project.

Supported in part by the National Science Foundation under Awards CCF-0430065, CCF-0635381, IIS-0915462, and CCF-1423560, in part by a gift from the Okawa Foundation, and in part by an Alfred P. Sloan Research Fellowship.