$2$ Biomedical applications of machine learning and computational modeling

Name $2$ examples of biomedical applications in which you might use a binary classifier; contrast each of those two examples to an example of regression in a similar biomedical application.

$3$ Nearest neighbor binary classification

Suppose we are developing a text reader for the visually impaired. We will start off with detecting numbers, in particular we will try to detect the numeral $"3".$ In this exercise you will use k$-$nearest neighbors to classify whether a digit is a $"3"$ or not.

$1.$ Load the UCI ML handwritten digits dataset from sklearn $($using load$\_$digits$).$

$2.$ Examine the data. e$.$g$.,$ determine how are the handwritten numerals stored? Use imshow to view a random sample of $10$ handwritten digits. Understand what the targets are for this dataset and again how they are stored.

$3.$ Using a $\backslash (80/20\backslash \%\backslash )$ split, train a k$-$nearest neighbors classifier to correctly classify $"3"$s$,$ reserving $\backslash (20\backslash \%\backslash )$ for the testing. First split the data, reserving $\backslash (20\backslash \%\backslash )$ for testing. Use a random$\_$state of $211.$

$4.$ Now create the appropriate binary target array.

$5.$ Plot a sample of digits $($say the first $40)$ along with the target value above it$.$

$6.$ Now try applying your classifier to the test data. Show the prediction for the same samples you showed above.

How accurate was your classifier? Show the confusion matrix for your classifier.

Try KNN for different numbers of neighbors $1$ through $7.$ How does the accuracy change? Show

the true positive rate and false positive rate. How does this compare with your expectation

based on your understanding of the KNN algorithm? If it does not match your expectation,

comment on some possible explanations.

$4$ K$-$means classifier

In this section, you will apply the K$-$means classifier to determine whether a given digit is a $3$ or

not a $3.$

How many classes are there in this problem? Let's define this as nClass

Apply K$-$means using nClass clusters.

Compare the performance of the binary KNN classifier with a binary K$-$means classifier.

Now try K$-$means using $2,3,$dots,$10,11($from $2$ through $11)$ clusters. Compare the correct

classification rate for each k value. Which seems to be the optimal number of clusters?

Explain with code