Name

Id No$.$

Branch

Mechanical

Note: Attempt all questions. Q$\mathrm{.}1-16$ carry $1$ mark each and Q$17$&$\mathrm{.}18$ carry $2$ marks each. Give the answer by encircling the correct

option $($a$),($b$),($c$)$ or $($d$)$ or by writing as directed in question. Multiple answers for same question or changing the option by

cutting will be treated as wrong answer and no marks will be awarded.

$1$ S$1$: Perceptron is a two layer network and weights of both layers' neurons are trainable

S$2$: The input applied to first layer lies in image space and its output forms pattern space.

S$3$: It is necessary to make the input patterns to be linear separable before applying to second layer.

S$4$: The solution obtained on convergence by each perceptron is always unique.

True Statements $($out of $S\frac{1}{S}\frac{2}{S}\frac{3}{S}4)$ are$-$

$2$ The XOR problem is said to be a benchmark problem because

$($a$)$ it can be dealt with a single node network

$($b$)$ It is a test problem with linearly non separable patterns

$($c$)$ It is a binary classification problem

$($d$)$ It is the simplest problem for binary classification

$3$ If the capacity of a perceptron to classify pattern is $8.$ The number of nodes in first layer of this perceptron is

$4$ S$1$: Perceptron was proposed by Rosenblatt as a two$-$layer feed$-$forward network of threshold logic units.

S$2$: Rosenblatt's perceptron evolved as merger of concepts proposed by McCulloch$-$Pitts and Hebb

False Statements $($out of $S\frac{1}{\text{\$}}2)$ are

$5$ While using fixed increment rule in perceptron learning algorithm, augmentation by appending $+1$ to pattern vector is to

$($a$)$ Convert pattern space to image space

$($b$)$ Convert image space to pattern space

$($c$)$ Provide the blas input to nodes

$($d$)$ Remove the effect of bias from nodes

$6$ S$1$: In method of steepest descent, weights are adjusted in the direction opposite to gradient vector.

S$2$: The convergence of steepest descent is guaranteed for any positive value of learning rate.

True Statements $($out of $S\frac{1}{S}2)$ are$-$

$7$ S$1$: The LMS algorithm and Widrow$-$Hoff Rule are not similar in nature.

S$2$: As LMS algorithm depends on estimate of gradient vector, it cannot be referred as stochastic gradient algorithm.

S$3$: Though LMS is faster than steepest descent in convergence but it is not robust and model Independent.

True Statements $($out of $S\frac{1}{\text{\$}}\frac{2}{\text{\$}}3)$ are

$8$ In the steepest descent algorithm the trajectory of weight vector in eight space is

while in case of an LMS

algorithm it is

$($a$)$ Well$-$defined, random

$($b$)$ Random, weil$-$defined

$($c$)$ Well defined in both

$($d$)$ Random in both

$9$ S$1$: The feedback loop around the weight vector in LMS algorithm behaves like a high$-$pass filter.

S$2$: The inverse of the learning$-$rate parameter is a measure of the memory of the LMS algorithm.

S$3$: The LMS algorithm operates with linear neuron.

False Statements $($out of $$\frac{1}{52}\frac{?}{53})$ are$-$

$10$ In multilayer perceptron

S$1$: If activation function of all nodes is nonlinear, then it will reduce to single layer perceptron

S$2$: hidden neurons enable the network to learn complex tasks by extracting meaningful features from input patterns

S$3$:the network exhibits a low degree of connectivity

False Statements $($out of $S\frac{1}{52}\frac{?}{53})$ are$-$

$11$ In Back$-$Propagation algorithm, the use of minus sign in delta rule accounts for

$($a$)$ Gradient descent in error space

$($b$)$ Gradient ascent in error space

$($c$)$ Gradient descent in weight space

$($d$)$ Gradient ascent in weight space