Computer Science 246/446

Homework

Spring 2016

  • No credit for late homework.
  • Programming assignments must be in Python, and must run on your instructional account on cycle1.csug.rochester.edu.
  • Use numpy for matrix operations.
  • To get an instructional account, see Marty Guenther in CSB 735.
  • To turn in programming assignments, run the command /u/cs246/TURN_IN

Homework 1, due in class Wed 1/20
  • Bishop 1.3
  • Bishop 1.11
  • Are the following properties true? Prove or disprove.
    1. (X ⊥ W|Z,Y) ∧ (X ⊥ Y|Z) ⇒ (X ⊥ Y,W|Z)
    2. (X ⊥ Y|Z) ∧ (X ⊥ Y|W) ⇒ (X ⊥ Y|Z,W)
Homework 2, due Fri 1/22 5pm by turnin script
Implement linear regression for the adult income dataset using Python. Data available in /u/cs246/data/adult and /u/cs446/data/adult. Take the sign of the output of linear regression as the predicted class label in order to compute the accuracy of your classifier.
Homework 3, due Mon 2/1 5pm by turn_in script
Implement naive bayes with Dirichlet smoothing for the adult income dataset. Plot classification error on the dev set as a function of alpha, and report the performance of the best alpha on the test set.
Homework 4, due Wed 2/10 in class
Bishop 4.8, 5.4, 5.7, 5.17
Homework 5, due Wed 2/17 5pm by turn_in script
Implement perceptron and SVM for the adult income dataset. Report how performance on dev varies as a function of C, and how perceptron and SVM compare.
Homework 6, due Wed 3/2 in class
Review problems handed out in class.
Homework 7, due Wed 3/30 5pm by turn_in script

Implement EM fitting of a mixture of gaussians on the two-dimensional data set points.dat. You should try different numbers of mixtures, as well as tied vs. separate covariance matrices for each gaussian. OR Implement EM fitting of the aspect model on the discrete data set pairs.dat. You should try different numbers of mixtures.

IN EITHER CASE Use the final 1/10 of the data for dev. Plot likelihood on train and dev vs iteration for different numbers of mixtures.

Homework 8, due Fri 4/15 5pm by turn_in script
Implement EM to train an HMM for whichever dataset you used for assignment 7. The observation probs should be as in assignment 7: either gaussian, or two discrete distributions conditionally independent given the hidden state. Does the HMM model the data better than the original non-sequence model? What is the best number of states?
Homework 9, due Wed 4/27 in class
Review problems handed out in class.

gildea @ cs rochester edu
April 19, 2016