CS485/685 - Assignments

There will be four assignments given the course, each worth 12.5% of the final mark (7.5% for CS685). Each assignment will have a theoretical part and a programming part.  Assignments are done individually (i.e., no team).  You are free to program in the language of your choice, however Matlab is recommended since it provides a convenient high-level programming environment for matrix operations.  If you decide to program in Matlab, the IST group maintains a nice set of online references for Matlab including a tutorial.

The approximate out and due dates are:

• A1: out Jan 10, due Jan 26 out Jan 17, due Feb 2
  
• A2: out Jan 26, due Feb 14 out Feb 2, due Feb 28
  
• A3: out Feb 14, due March 6 out March 1, due March 15
  
• A4: out March 6, due March 22 out March 14, due March 29
  

• Assignment handout
• Datasets: data.zip
  
• Problem: this data is a modified version of the Optical Recognition of Handwritten Digits Dataset from the UCI repository.  It contains preprocessed black and white images of the digits 5 and 6.  Each attribute indicates how many pixels are black in a patch of 4 x 4 pixels.
  
• Format: there is one row per image and one column per attribute.  The class labels are 5 and 6.
• Parsing: if you use Matlab, you can load the datasets easily with the command "load dataX.csv".
• Ricardo Salmon (rsalmon [at] cs [dot] uwaterloo [dot] ca) is the TA in charge of Assignment 1.  He will hold special office hours on Jan 30, 11 am - 1 pm in the AI lab (DC2306C).

• Assignment handout
• Corrected dataset: dataset.zip (Feb 13) make sure to download this corrected dataset
  
• Problem: this data corresponds to samples from a 2D surface that you can plot to visualize what the algorithms are doing.  The data is divided in two sets, one for training and one for testing.  Unlike assignment 1, is it is not necessary to use cross validation.  For this assignment sinply train with the training data and report the accuracy with respect to the test data.  
  
• Format: there is one row per data instance and one column per attribute.  The labels are real values.
• Parsing: if you use Matlab, you can load the datasets easily with the command "load filename.csv".
• Visualization: here is a 3D view of the data.  To generate this surface, type the following command in matlab "for i=1:128, f(fDataTrain(i,1),fDataTrain(i,2))=fLabelsTrain(i); f(fDataTest(i,1),fDataTest(i,2))=fLabelsTest(i); end, mesh(f);" and rotate the axes appropriately.
  
• What to hand in:
• print out of your code
• regularlized generalized linear regression:
• graph that shows the mean squared error as a function of the regularization weigtht lambda (0 to 4 in increments of 0.1) and the degree of the polynomial basis functions (degree 1, 2, 3 and 4).  
  
• discussion of the results and how the running time varies with the degree.
  
• Bayesian generalized linear regression (assume output noise is Gaussian with variance = 1):
• graph that shows the mean squared error as a function of the prior parameter lambda (0 to 4 in increments of 0.1) and the degree of the polynomial basis functions (degree 1, 2, 3 and 4).  Use a Gaussian prior: Pr(w) = k*exp(-0.5*lambda*w^Tw).
  
• discussion of the results and how they compare with regularized generalized linear regression
• Gaussian process regression (assume output noise is Gaussian with variance = 1):
• polynomial kernel: graph that shows mean squared error as a function of the degree (1 to 5 in increments of 1)
• Gaussian kernel: graph that shows mean squared error as a function of the sigma parameter of the gaussian kernel (1 to 6 in increments of 1)
• identity kernel: mean squared error
• discussion of the results and how the running time varies
• Neural network regression (use a step length of 0.0001 and perform 10,000 rounds of training where each round consists of sequential gradient descent with respect to each training point)
• graph that shows mean squared error as a function of the number of hidden units (1 to 6 in increments of 1)
• discussion of the results
• Ricardo Salmon (rsalmon [at] cs [dot] uwaterloo [dot] ca) is the TA in charge of Assignment 2.  He will hold special office hours on Feb 13, 11 am - 1 pm in the AI lab (DC2306C).

• Assignment handout
• Dataset: data.zip
  
• Problem: this dataset consists of sequences of hidden states and noisy observations.
  
• Format: each the data file consists of a sequence of noisy observations with one data instance per row.  Each label file consists of a sequence of hidden states (labeled 1, 2 or 3).
  
• Parsing: if you use Matlab, you can load the dataset easily with the command "load filename.csv".
• Ruth Urner (rurner [at] cs [dot] uwaterloo [dot] ca) is the TA in charge of Assignment 3. She will hold special office hours on March 12, 2-4 pm in the AI lab (DC2306C).

• Assignment handout
• Ruth Urner (rurner [at] cs [dot] uwaterloo [dot] ca) is the TA in charge of Assignment 4. She will hold special office hours on March 26, 1-3 pm in the AI lab (DC2306C).