Sample 3

Question 1

For an N class classification problem, how many SVMs are required to implement and how many to consult when classifying a sample using each method:
1-1
1 all
Binary decision tree
Binary code

Answer 1

1-1: implement - N(N-1)/2 - consult - N(N-1)/2
1 vs All: implement - N - consult - N
Binary Decision Tree: Implement - N - 1 - consult ceiling(log_2(N))
Binary code: Implement - ceiling(log_2(N)) - consult - ceiling(log_2(N))

Question 2

Techniques to pretty overfitting on CNN

Answer 2

Dropout
Early stopping
Batch normalisation
Class preserving data augmentation

Question 3

Hard and soft classifiers for SVMs

Answer 3

Hard: sgn(z) =
-1 if z < 0
+1 if z >= 0

Soft: h(z) =
-1 if z < -1
z if -1 <= z <= 1
+1 if z > 1

Where z is discriminant & varies but for linear linearly separable case is w^Tx + w0

Question 4

Condition to check if point for linear SVM is classified correctly

Answer 4

Let y_i( w^Tx_i + w₀) be our key quantity k
If k >= 1, sample is correctly classified
If 0 <= k <1, sample is correctly classified but within band
If k < 0, sample is misclassified