Sample Exam 2

Question 1

Gradient update rule for multi class perception learning algorithm

Answer 1

a_predicted = a_predicted - η(y)
a_actual = a_actual + η(y)

Question 2

Rules to memorise for linear non linearly separable SVM

Answer 2

y_i(w^Tx_i)+ w₀
^ this quantity is
>= 1 for samples outside the band and correctly classified
Between 0 and 1, for samples inside the band and correctly classified
Less than 0 for misclassified

Question 3

Optimal discriminator

Answer 3

p_data(x)/ (p_data(x) + p_gen(x))

=

1 - p_gen(x)/(p_data(x) + p_gen(x))

Question 4

The threshold for gradient clipping is

Answer 4

A hyper parameter
Setting it too low could restrict a network’s ability to learn

Question 5

Ideal design for output layer of multi class classifier

Answer 5

One neuron per class, softmax activation

Question 6

Linear Discriminant analysis protects data onto

Answer 6

Directions in which the random different classes is most distinct