BYU CS 474 DEEP LEARNING MIDTERM STUDY GUIDE
FREQUENTLY TESTED ACTUAL EXAM QUESTIONS WITH
CORRECT ANSWERS
Question 1.
1. Increasing regularization _____ bias and ____ variance
ANSWER
a. Increases bias, reduces variance.
Question 2.
1. Dropout
ANSWER
L2 (Ridge) regularization
(set certain activations to zero/squared error penalty)
Question 3.
What does V(s_t) seek to approximate in the PPO algorithm
ANSWER
The expected discounted sum of rewards starting from the current state and continuing under
the current policy.
Question 4.
Vanilla transformer is a(n) ____ transformer variant architecture
ANSWER
encoder - decoder
Question 5.
the objective function for autoregressive generative models
ANSWER
Maximum likelihood
1/7
, Question 6.
GANs can tractably do ______
ANSWER
sampling
Question 7.
solutions for a small dataset
ANSWER
Data augmentation
Use an architecture with an inductive bias tailored to the dataset
Fine-tune a pre-trained model
Question 8.
Deep learning is particularly well-suited (relative to classical ML) to
ANSWER
i. Large dataset
ii. High-dimensional data
iii. Variable-length data
Question 9.
Flow-based generative models can tractably do ______
ANSWER
density estimation, sampling
Question 10.
1. In the skip-gram negative sampling algorithm, what would happen if you omitted negative
sampling?
ANSWER
All word embeddings would be learned to be the same.
2/7
FREQUENTLY TESTED ACTUAL EXAM QUESTIONS WITH
CORRECT ANSWERS
Question 1.
1. Increasing regularization _____ bias and ____ variance
ANSWER
a. Increases bias, reduces variance.
Question 2.
1. Dropout
ANSWER
L2 (Ridge) regularization
(set certain activations to zero/squared error penalty)
Question 3.
What does V(s_t) seek to approximate in the PPO algorithm
ANSWER
The expected discounted sum of rewards starting from the current state and continuing under
the current policy.
Question 4.
Vanilla transformer is a(n) ____ transformer variant architecture
ANSWER
encoder - decoder
Question 5.
the objective function for autoregressive generative models
ANSWER
Maximum likelihood
1/7
, Question 6.
GANs can tractably do ______
ANSWER
sampling
Question 7.
solutions for a small dataset
ANSWER
Data augmentation
Use an architecture with an inductive bias tailored to the dataset
Fine-tune a pre-trained model
Question 8.
Deep learning is particularly well-suited (relative to classical ML) to
ANSWER
i. Large dataset
ii. High-dimensional data
iii. Variable-length data
Question 9.
Flow-based generative models can tractably do ______
ANSWER
density estimation, sampling
Question 10.
1. In the skip-gram negative sampling algorithm, what would happen if you omitted negative
sampling?
ANSWER
All word embeddings would be learned to be the same.
2/7