Georgia Tech CS 7643 Deep Learning Quiz 1 Practice | 30 Original Questions, Answers &
Detailed Rationales | Fall 2026.
Section I — Machine Learning & Parametric
Models
Question 1
Which statement best characterizes a parametric model?
A. The number of stored parameters must increase with every additional training example.
B. Once the model architecture is specified, learning estimates a finite set of parameters.
C. The model must store the entire training set to make predictions.
D. Parametric models cannot represent nonlinear functions.
Answer
B. Once the model architecture is specified, learning estimates a finite set of parameters.
Rationale
A parametric model assumes a particular functional form characterized by a finite-dimensional
parameter vector.
For example, a linear classifier
f(x)=Wx+bf(x)=Wx+b
has parameters WW and bb. Adding more training examples does not automatically increase the
dimensions of WW or bb.
K-nearest neighbors is a classic contrast because prediction depends directly on stored training
examples.
Question 2
Suppose a multiclass linear classifier receives
x∈Rdx\in\mathbb{R}^{d}
,and produces scores for KK classes using
s=Wx+b.s=Wx+b.
What should the dimensions of WW and bb be?
A. W∈Rd×KW\in\mathbb{R}^{d\times K}, b∈Rdb\in\mathbb{R}^{d}
B. W∈RK×dW\in\mathbb{R}^{K\times d}, b∈RKb\in\mathbb{R}^{K}
C. W∈RK×KW\in\mathbb{R}^{K\times K}, b∈Rdb\in\mathbb{R}^{d}
D. W∈Rd×dW\in\mathbb{R}^{d\times d}, b∈RKb\in\mathbb{R}^{K}
Answer
B.
W∈RK×d,b∈RKW\in\mathbb{R}^{K\times d}, \qquad b\in\mathbb{R}^{K}
Rationale
Multiplication requires
(K×d)(d×1)=K×1.(K\times d)(d\times1)=K\times1.
Therefore WxWx produces one score for each of the KK classes. The bias vector must
consequently contain KK entries.
Question 3
Consider two classes with linear scores
s1(x)=w1Tx+b1s_1(x)=w_1^Tx+b_1
and
s2(x)=w2Tx+b2.s_2(x)=w_2^Tx+b_2.
Which equation defines their decision boundary?
A.
w1Tx+b1=0w_1^Tx+b_1=0
B.
, w2Tx+b2=0w_2^Tx+b_2=0
C.
(w1−w2)Tx+(b1−b2)=0(w_1-w_2)^Tx+(b_1-b_2)=0
D.
(w1+w2)Tx+(b1+b2)=1(w_1+w_2)^Tx+(b_1+b_2)=1
Answer
C.
(w1−w2)Tx+(b1−b2)=0\boxed{(w_1-w_2)^Tx+(b_1-b_2)=0}
Rationale
The boundary occurs where the classifier assigns equal scores:
s1(x)=s2(x).s_1(x)=s_2(x).
Therefore,
w1Tx+b1=w2Tx+b2.w_1^Tx+b_1=w_2^Tx+b_2.
Rearranging gives
(w1−w2)Tx+(b1−b2)=0.(w_1-w_2)^Tx+(b_1-b_2)=0.
This is a hyperplane.
Question 4
A model produces logits
z=(2,5,−1).z=(2,5,-1).
If the same constant c=10c=10 is added to every logit, what happens to the softmax
probabilities?
A. Every probability increases.
B. Every probability decreases.
Detailed Rationales | Fall 2026.
Section I — Machine Learning & Parametric
Models
Question 1
Which statement best characterizes a parametric model?
A. The number of stored parameters must increase with every additional training example.
B. Once the model architecture is specified, learning estimates a finite set of parameters.
C. The model must store the entire training set to make predictions.
D. Parametric models cannot represent nonlinear functions.
Answer
B. Once the model architecture is specified, learning estimates a finite set of parameters.
Rationale
A parametric model assumes a particular functional form characterized by a finite-dimensional
parameter vector.
For example, a linear classifier
f(x)=Wx+bf(x)=Wx+b
has parameters WW and bb. Adding more training examples does not automatically increase the
dimensions of WW or bb.
K-nearest neighbors is a classic contrast because prediction depends directly on stored training
examples.
Question 2
Suppose a multiclass linear classifier receives
x∈Rdx\in\mathbb{R}^{d}
,and produces scores for KK classes using
s=Wx+b.s=Wx+b.
What should the dimensions of WW and bb be?
A. W∈Rd×KW\in\mathbb{R}^{d\times K}, b∈Rdb\in\mathbb{R}^{d}
B. W∈RK×dW\in\mathbb{R}^{K\times d}, b∈RKb\in\mathbb{R}^{K}
C. W∈RK×KW\in\mathbb{R}^{K\times K}, b∈Rdb\in\mathbb{R}^{d}
D. W∈Rd×dW\in\mathbb{R}^{d\times d}, b∈RKb\in\mathbb{R}^{K}
Answer
B.
W∈RK×d,b∈RKW\in\mathbb{R}^{K\times d}, \qquad b\in\mathbb{R}^{K}
Rationale
Multiplication requires
(K×d)(d×1)=K×1.(K\times d)(d\times1)=K\times1.
Therefore WxWx produces one score for each of the KK classes. The bias vector must
consequently contain KK entries.
Question 3
Consider two classes with linear scores
s1(x)=w1Tx+b1s_1(x)=w_1^Tx+b_1
and
s2(x)=w2Tx+b2.s_2(x)=w_2^Tx+b_2.
Which equation defines their decision boundary?
A.
w1Tx+b1=0w_1^Tx+b_1=0
B.
, w2Tx+b2=0w_2^Tx+b_2=0
C.
(w1−w2)Tx+(b1−b2)=0(w_1-w_2)^Tx+(b_1-b_2)=0
D.
(w1+w2)Tx+(b1+b2)=1(w_1+w_2)^Tx+(b_1+b_2)=1
Answer
C.
(w1−w2)Tx+(b1−b2)=0\boxed{(w_1-w_2)^Tx+(b_1-b_2)=0}
Rationale
The boundary occurs where the classifier assigns equal scores:
s1(x)=s2(x).s_1(x)=s_2(x).
Therefore,
w1Tx+b1=w2Tx+b2.w_1^Tx+b_1=w_2^Tx+b_2.
Rearranging gives
(w1−w2)Tx+(b1−b2)=0.(w_1-w_2)^Tx+(b_1-b_2)=0.
This is a hyperplane.
Question 4
A model produces logits
z=(2,5,−1).z=(2,5,-1).
If the same constant c=10c=10 is added to every logit, what happens to the softmax
probabilities?
A. Every probability increases.
B. Every probability decreases.
