Description
1. Run the kNN code with k=2. Display some of the misclassified digits
alongside their nearest neighbours. Can you think of any reason why
these digits might be more likely to be misclassified? Is there anything
they have in common? Does this suggest any way in which the algorithm
might be improved?
2. Try to display two of the classes (e.g. 8s and 3s) in 3D space using the 3D
plotting code. Edit the code so that the misclassified digits stand out in
the 3D space with respect to the other digits. You might have to use a
different plotting symbol to do this.
Do the misclassified digits lie in a particular region? Is there anything
different about their location with respect to the main clusters? Does
this suggest any way in which the algorithm might be improved?
3. Find the accuracy of the kNN algorithm with different numbers of
eigenvectors, where the number n could range from 2 up to 256 in steps
of 5 or 10.
What is the shape of the curve? Does it increase steadily? Does it level
off or does it have a peak? Can you explain the shape?
How does the execution time change as the number of eigenvectors
changes? Is there an optimum number?
4. Try to run the code which selects digits within a box. Try to change the
size of the box? How does the accuracy change? How does the execution
time change?
5. Try to run the code which uses blurring to selects digits within a box. Try
to change the size of the box? How does the accuracy change? How does
the execution time change?
Try changing the level of blurring? What effect does it have on accuracy?
Can you come up with a multi-level strategy with different levels of
blurring and different box sizes? Can you improve both the accuracy and
execution time?