Project 6: Classification

Which Digit?
Which are Faces?

Due 5/08 at 11:59pm


In this project, you will design three classifiers: a naive Bayes classifier and a perceptron classifier and a large-margin (MIRA) classifier. You will test your classifiers on two image datasets: a set of scanned handwritten digit images and a set of face images in which edges have already been detected. Even your simple classifiers will be able to do quite well on these tasks when given enough training data.

Optical character recognition (OCR) is the task of extracting text from image sources. The first data set on which you will run your classifiers is a collection of handwritten numerical digits (0-9). This is a very commercially useful technology, similar to the technique used by the US post office to route mail by zip codes. There are systems that can perform with over 99% classification accuracy (see LeNet-5 for an example system in action).

Face detection is the task of localizing faces within video or still images. The faces can be at any location and vary in size. There are many applications for face detection, including human computer interaction and surveillance applications. You will attempt a reduced face detection task in which your system is presented with an image that has been pre-processed by an edge detection algorithm. The task is to determine whether the edge image is a face or not. There are several systems in use that perform quite well at the face detection task. One good system is the Face Detector by Schneiderman and Kanade. You can even try it out on your own photos in this demo.

The code for this project spans the following files and data files, available as a zip file

Data file Data file, including the digit and face data.
Files you will edit The location where you will write your naive Bayes classifier. The location where you will write your perceptron classifier. The location where you will write your MIRA classifier. The wrapper code that will call your classifiers. You will also write your enhanced feature extractor here. You will also use this code to analyze the behavior of your classifier.
Files you should read but NOT edit Abstract super class for the classifiers you will write.
(You should read this file carefully to see how the infrastructure is set up.) I/O code to read in the classification data. Code defining some useful tools. You may be familiar with some of these by now, and they will save you a lot of time. A simple baseline classifier that just labels every instance as the most frequent class.

What to submit: You will fill in portions of,, and (only) during the assignment, and submit them.

Evaluation: Your code will be autograded for technical correctness. Please do not change the names of any provided functions or classes within the code, or you will wreak havoc on the autograder.

Academic Dishonesty: We will be checking your code against other submissions in the class for logical redundancy. If you copy someone else's code and submit it with minor changes, we will know. These cheat detectors are quite hard to fool, so please don't try. We trust you all to submit your own work only; please don't let us down. Instead, contact the course staff if you are having trouble.

Getting Started

To try out the classification pipeline, run from the command line. This will classify the digit data using the default classifier (mostFrequent) which blindly classifies every example with the most frequent label.


As usual, you can learn more about the possible command line options by running:

python -h  

We have defined some simple features for you. Later you will implement more intelligent features. Our simple features have one feature for each pixel location, which can take values 0 or 1. The features are encoded as a Counter where keys are feature locations (represented as (column,row)) and values are 0 or 1. The face recognition data set has value 1 only for those pixels identified by a Canny edge detector.

Naive Bayes

A skeleton implementation of a naive Bayes classifier is provided for you in You will fill in the trainAndTune function, the calculateLogJointProbabilities function and the findHighOddsFeatures function.


A naive Bayes classifier models a joint distribution over a label $Y$ and a set of observed random variables, or features, $\{F_1, F_2, \ldots F_n\}$, using the assumption that the full joint distribution can be factored as follows (features are conditionally independent given the label):

P(F_1 \ldots F_n, Y) = P(Y) \prod_i P(F_i \vert Y)

To classify a datum, we can find the most probable label given the feature values for each pixel, using Bayes theorem:

P(y \vert f_1, \ldots, f_m) &=& \frac{P(f_1, \ldots, f_m \...
&=& \textmd{arg max}_{y} P(y) \prod_{i = 1}^m P(f_i \vert y)

Because multiplying many probabilities together often results in underflow, we will instead compute log probabilities which have the same argmax:

\textmd{arg max}_{y} log(P(y \vert f_1, \ldots, f_m) &=& \te...
...{arg max}_{y} (log(P(y)) + \sum_{i = 1}^m log(P(f_i \vert y)))

To compute logarithms, use math.log(), a built-in Python function.

Parameter Estimation

Our naive Bayes model has several parameters to estimate. One parameter is the prior distribution over labels (digits, or face/not-face), $P(Y)$.

We can estimate $P(Y)$ directly from the training data:

\hat{P}(y) = \frac{c(y)}{n}

where $c(y)$ is the number of training instances with label y and n is the total number of training instances.

The other parameters to estimate are the conditional probabilities of our features given each label y: $P(F_i \vert Y = y)$. We do this for each possible feature value ($f_i \in {0,1}$).

\hat{P}(F_i=f_i\vert Y=y) &=& \frac{c(f_i,y)}{\sum_{f_i}{c(f_i,y)}} \\

where $c(f_i,y)$ is the number of times pixel $F_i$ took value $f_i$ in the training examples of label y.


Your current parameter estimates are unsmoothed, that is, you are using the empirical estimates for the parameters $P(f_i\vert y)$. These estimates are rarely adequate in real systems. Minimally, we need to make sure that no parameter ever receives an estimate of zero, but good smoothing can boost accuracy quite a bit by reducing overfitting.

In this project, we use Laplace smoothing, which adds k counts to every possible observation value:

$P(F_i=f_i\vert Y=y) = \frac{c(F_i=f_i,Y=y)+k}{\sum_{f_i}{(c(F_i=f_i,Y=y)+k)}}$

If k=0, the probabilities are unsmoothed. As k grows larger, the probabilities are smoothed more and more. You can use your validation set to determine a good value for k. Note: don't smooth P(Y).

Question 1 (6 points) Implement trainAndTune and calculateLogJointProbabilities in In trainAndTune, estimate conditional probabilities from the training data for each possible value of k given in the list kgrid. Evaluate accuracy on the held-out validation set for each k and choose the value with the highest validation accuracy. In case of ties, prefer the lowest value of k. Test your classifier with:

python -c naiveBayes --autotune 

Hints and observations:

Odds Ratios

One important tool in using classifiers in real domains is being able to inspect what they have learned. One way to inspect a naive Bayes model is to look at the most likely features for a given label.

Another, better, tool for understanding the parameters is to look at odds ratios. For each pixel feature $F_i$ and classes $y_1, y_2$, consider the odds ratio:

\mbox{odds}(F_i=on, y_1, y_2) = \frac{P(F_i=on\vert y_1)}{P(F_i=on\vert y_2)}

This ratio will be greater than one for features which cause belief in $y_1$ to increase relative to $y_2$.

The features that have the greatest impact at classification time are those with both a high probability (because they appear often in the data) and a high odds ratio (because they strongly bias one label versus another).

Question 2 (2 points) Fill in the function findHighOddsFeatures(self, label1, label2). It should return a list of the 100 features with highest odds ratios for label1 over label2. The option -o activates an odds ratio analysis. Use the options -1 label1 -2 label2 to specify which labels to compare. Running the following command will show you the 100 pixels that best distinguish between a 3 and a 6.

python -a -d digits -c naiveBayes -o -1 3 -2 6  


A skeleton implementation of a perceptron classifier is provided for you in You will fill in the train function, and the findHighOddsFeatures function.

Unlike the naive Bayes classifier, a perceptron does not use probabilities to make its decisions. Instead, it keeps a weight vector $w^y$ of each class $y$ ($y$ is an identifier, not an exponent). Given a feature list $f$, the perceptron compute the class $y$ whose weight vector is most similar to the input vector $f$. Formally, given a feature vector $f$ (in our case, a map from pixel locations to indicators of whether they are on), we score each class with:

\mbox{score}(f,y) = \sum_i f_i w^y_i
Then we choose the class with highest score as the predicted label for that data instance. In the code, we will represent $w^y$ as a Counter.

Learning weights

In the basic multi-class perceptron, we scan over the data, one instance at a time. When we come to an instance $(f, y)$, we find the label with highest score:
y' = \textmd{arg max}_{y''} score(f,y'')

We compare $y'$ to the true label $y$. If $y' = y$, we've gotten the instance correct, and we do nothing. Otherwise, we guessed $y'$ but we should have guessed $y$. That means that $w^y$ should have scored $f$ higher, and $w^{y'}$ should have scored $f$ lower, in order to prevent this error in the future. We update these two weight vectors accordingly:

w^y += f
w^{y'} -= f

Using the addition, subtraction, and multiplication functionality of the Counter class in, the perceptron updates should be relatively easy to code. Certain implementation issues have been taken care of for you in, such as handling iterations over the training data and ordering the update trials. Furthermore, the code sets up the weights data structure for you. Each legal label needs its own Counter full of weights.

Question 3 (4 points) Fill in the train method in Run your code with:

python -c perceptron 

Hints and observations:

Visualizing weights

Perceptron classifiers, and other discriminative methods, are often criticized because the parameters they learn are hard to interpret. To see a demonstration of this issue, we can repeat the visualization exercise from the naive Bayes classifier.

Question 4 (1 point) Fill in findHighOddsFeatures(self, label1, label2) in It should return a list of the 100 features with highest difference in feature weights. You can display the 100 pixels with the largest difference in weights $w^3 - w^6$ using the command:

python -c perceptron -o -1 3 -2 6  


A skeleton implementation of the MIRA classifier is provided for you in MIRA is an online learner which is closely related to both the support vector machine and perceptron classifiers. You will fill in the trainAndTune function.


Similar to a multi-class perceptron classifier, multi-class MIRA classifier also keeps a weight vector $w^y$ of each label $y$. We also scan over the data, one instance at a time. When we come to an instance $(f, y)$, we find the label with highest score:

We compare $y'$ to the true label $y$. If $y' = y$, we've gotten the instance correct, and we do nothing. Otherwise, we guessed $y'$ but we should have guessed $y$. Unlike perceptron, we update the weight vectors of these labels with variable step size:

where is chosen such that it minimizes
subject to the condition that

which is equivalent to
subject to and

Note that, , so the condition is always true given Solving this simple problem, we then have
However, we would like to cap the maximum possible value of by a positive constant C, which leads us to

Question 5 (6 points) Implement trainAndTune in This method should train a MIRA classifier using each value of C in Cgrid. Evaluate accuracy on the held-out validation set for each C and choose the C with the highest validation accuracy. In case of ties, prefer the lowest value of C. Test your MIRA implementation with:
python -c mira --autotune 

Hints and observations:

Feature Design

Building classifiers is only a small part of getting a good system working for a task. Indeed, the main difference between a good classification system and a bad one is usually not the classifier itself (e.g. perceptron vs. naive Bayes), but rather the quality of the features used. So far, we have used the simplest possible features: the identity of each pixel (being on/off).

To increase your classifier's accuracy further, you will need to extract more useful features from the data. The EnhancedFeatureExtractorDigit in is your new playground. When analyzing your classifiers' results, you should look at some of your errors and look for characteristics of the input that would give the classifier useful information about the label. You can add code to the analysis function in to inspect what your classifier is doing. For instance in the digit data, consider the number of separate, connected regions of white pixels, which varies by digit type. 1, 2, 3, 5, 7 tend to have one contiguous region of white space while the loops in 6, 8, 9 create more. The number of white regions in a 4 depends on the writer. This is an example of a feature that is not directly available to the classifier from the per-pixel information. If your feature extractor adds new features that encode these properties, the classifier will be able exploit them. Note that some features may require non-trivial computation to extract, so write efficient and correct code.

Question 6 (6 points) Add new features for the digit dataset in the EnhancedFeatureExtractorDigit function in such a way that it works with your implementation of the naive Bayes classifier: this means that for this part, you are restricted to features which can take a finite number of discrete values (and if you have assumed that features are binary valued, then you are restricted to binary features). Note that you can encode a feature which takes 3 values [1,2,3] by using 3 binary features, of which only one is on at the time, to indicate which of the three possibilities you have. In theory, features aren't conditionally independent as naive Bayes requires, but your classifier can still work well in practice. We will test your classifier with the following command:

python -d digits -c naiveBayes -f -a -t 1000  
With the basic features (without the -f option), your optimal choice of smoothing parameter should yield 82% on the validation set with a test performance of 78%. You will receive 3 points for implementing new feature(s) which yield any improvement at all. You will receive 3 additional points if your new feature(s) give you a test performance greater than or equal to 85% with the above command.

Congratulations! You're finished with the CS 188 projects.