Description
Perceptron algorithm for Optical Character Recognition
(total points: 80 pts + 10 report pts + 10 result pts)
Task description. In Optical Character Recognition (OCR) we seek to predict a number (between 0 to
9) for a given image of handwritten digit. In this assignment we consider a simplified binary classification
task: differentiating two digits 3 versus 5 and you will implement and experiment with variations of the
perceptron algorithm.
Data. All the handwritten digits are originally taken from http://www.kaggle.com/c/digit-recognizer/data
The original dataset contains the sample digits suitable for OCR. We extract samples with only labels 3 and
5. Following a little pre-processings we produce three datasets for this assignment as follows:
(a) Train Set (pa2 train.csv): Includes 4888 rows (samples). Each sample is in fact a list of 785 values.
The first number is the digit’s label which is 3 or 5. The other 784 floating values are the the flattened
gray-scale values from a 2d digital handwritten image with shape 28 × 28.
(b) Validation Set (pa2 valid.csv): Includes 1629 rows. Each row obeys the same format given for the
train set. This set will be used to select your best trained model.
(c) Test Set (pa2 test.csv): Includes 1629 rows. Each row contains only 784 numbers. The label column
is omitted from each row.
Important Guidelines. For all parts of this assignment:
(a) Please assign labels +1 to number 3 and -1 to label 5. In your produced predictions, please use only
+1 and -1 as labels not 3 and 5.
(b) Please do not shuffle the given data unless instructed to do so. While in practice shuffling should be
used to improve training convergence, for this assignment we ask you not to shuffle the data to ensure
deterministic results for assessment purpose.
(c) To simplify the notation in this assignment, your load function which loads train, validation and test
dataset should add a bias feature to the dataset. The bias feature is a feature with value of 1.0 for all
of the samples. Therefore the feature size for each samples will become 785.
2
Part 1 (20 pts) : Online Perceptron. In the online perceptron algorithm we train a linear classifier
with parameter w to predict the label of a sample with equation:
yˆ = sign(wT x) (1)
Where ˆy ∈ {−1, 1}. Algorithm 1 describes the online perceptron.
Algorithm 1 Online Perceptron
1: procedure OnlinePerceptron
2: w ← 0
3: while iter < iters:
4: for all sample xt in train set: // no shuffling
5: if ytwT xt ≤ 0:
6: w ← w + ytxt
7: return w
In this part we are interested in the following experiments for the online perceptron algorithm:
(a) Implement online perceptron as described in Algorithm 1. Set the iters = 15. During the training,
at the end of each iteration use the current w to make prediction on the validation samples. Record
the accuracies for the train and validation at the end of each iteration. Plot the recorded train and
validation accuracies versus the iteration number. Does the train accuracy reach to 100%? Why?
(b) Use the validation accuracy to decide the best value for iters. Apply the corresponding learned model
to make predictions for the samples in the test set. Generate the prediction file oplabel.csv. Please
note that your file should only contain +1 (for 3) and -1 (for 5) and the number of rows should be the
same as pa2 test.csv.
Part 2 (20 pts) : Average Perceptron. In this part you will implement and experiment with average
perceptron, which is described in Algorithm 2.
Algorithm 2 Average Perceptron
1: procedure AveragePerceptron
2: w ← 0 //initialize weight vector
3: w¯ ← 0 // initialize average weight vector
4: s ← 1 //initialize the example counter to 1
5: while iter < iters:
6: for all sample xt in the train set: // no shuffling
7: if ytwT xt ≤ 0:
8: w ← w + ytxt
9: w¯ ← sw¯+w
s+1 // update w¯ regardless of w udpate
10: s ← s + 1
11: return w¯
Note that in Algorithm 2, the running average w¯ always gets updated every time an example is process,
regardless whether the current w correctly classify it or not 1
.
Perform the following experiments:
(a) Apply your implemeted average perceptron to learn from the training data. Plot the train and validation accuracies versus the iteration number for iters = 1, …, 15.
1
If you are interested, there is a slightly different but equivalent version of this algorithm that only performs update w¯ when
w is updated, presented in Chapter 3 of A course in machine learning (Algorithm 7). You can choose to implement either
version.
3
(b) What are your observations when comparing the training accuracy and validation accuracy curves
of the average perceptron with those of the online perceptron? What are your explanation for the
observations?
(c) Use the validation accuracy to decide the best value for iter and apply the corresponding learned model
to make predictions for the test data. Please name the predicted file as aplabel.csv.
Part 3 (40 pts). Polynomial Kernel Perceptron. The online/average perceptron in Algorithm( 1
and 2) are linear models. In this part we will consider kernelized perceptron, as described in Algorithm 3,
with a polynomial kernel kp of degree p:
kp(x1, x2) = (1 + x
T
1 x2)
p
(2)
Algorithm 3 Kernel (polynomial) Perceptron
1: procedure KernelPerceptron
2: αi ← 0 for i = 1, …, N
3: compute the gram matrix K(i, j) = kp(xi
, xj )
4: while iter < iters:
5: for all sample xt in training set: // no shuffling
6: u =
P
i αiK(i, t)yi
7: if ytu ≤ 0:
8: αt ← αt + 1
Implement the kernelized perceptron with polynomial kernel and perform the following experiments:
(a) Apply the kernelized perceptron with different p values in [1, 2, 3, 4, 5]:
1) For each p value, run your algorithm with iter = 15. At the end of each iteration use the current
model (aka α’s) to make prediction for the validation set. Record the train and validation accuracy
for each iteration and plot the train and validation accuracies versus the iteration number.
2) Record the best validation accuracy achieved for each p over all iterations. Plot the recorded best
validation accuracies versus p. How do you think p is affecting the train and validation performance?
(b) Use your best model (the best one you found over all p values and all iterations above) to make
prediction for the test set. Please name the predicted file as kplabel.csv.
Submission. Your submission should include the following:
1) Your source code with a short instruction on how to run the code in a readme.txt.
2) Your report only in pdf, which begins with a general introduction section, followed by one section for each
part of the assignment.
3) Three prediction files oplabel.csv, aplabel.csv and kplabel.csv. These prediction file will be scored against
the ground truth y values and 10% of the grade will be based on this score.
4) Please note that all the files should be in one folder and compressed only by .zip.
4
