CSE 493 G1/ 599 G1 Deep Learning Spring 2024 Quiz 2 - Solutions

Sep 30, 2026

Deep Learning course quiz 2 solutions for CSE 493 G1 / 599 G1, Spring 2024, dated Feb 2, 2024. The document provides an Honor Code acknowledgment and a structured answer key covering 20-minute double-sided exam content. It includes scoring guidance and full solution explanations for five True/False items on ReLU dead neurons, layer normalization behavior, batch normalization placement, and dropout applicability, plus a multi-choice section addressing overfitting, activation functions, and correct option sets.

LanguageEnglish

CategoryExam

FormatPDF

Pages8

Reading Time20 min

AccessFree to download

Quiz 2 Instructions

True/False Questions

1.1 ReLU and dead neurons

1.2 Layer normalization training/testing behavior

1.3 Batch normalization placement

1.4 Dropout layer restrictions

1.5 Saliency map purpose

Multiple Choice Questions

2.1 Overfitting diagnosis and remedies

2.2 Valid activation functions

CSE 493 G1/ 599 G1 Deep Learning Spring 2024 Quiz 2 SOLUTIONS Feb 2, 2024 Full Name: UW Net ID: | Question | Score | | --- | --- | | True/False (5 pts) | | | --- | --- | | Multiple Choice (8 pts) | | | Short Answer (9 pts) | | | Total (22 pts) | | Welcome to the CSE 493 G1 Quiz 2! • The exam is 20 min and is double-sided. • No electronic devices are allowed. I understand and agree to uphold the University of Washington Honor Code during this exam. Signature: Date: Good luck! This page is left blank for scratch work only. DO NOT write your answers here. 1 True / False (5 points) -Recommended 5 Minutes Fill in the circle next to True or False, or fill in neither. Fill it in completely like this: . No explanations are required. Scoring: Correct answer is worth 2 points. To discourage guessing, incorrect answers are worth-1 points. Leaving a question blank will give 0 points. 1.1 ReLU (Rectified Linear Unit) activation functions can suffer from the problem of dead neurons. \# True \# False SOLUTION: True, In a ReLU activation, if the input is less than zero, the output is zero. This can lead to situations where neurons stop responding to inputs altogether, known as dead ReLU problem. 1.2 One downside of layer normalization is that it works differently during training and testing, leading toa common source of errors. \# True \# False SOLUTION: False. Layer normalization maintains consistent behavior between training and testing, as it computes normalization statistics from the inputs to each layer for each training case, unlike batch normalization which behaves differently in these phases. 1.3 Batch normalization is applied to the input of each layer before the activation function. \# True \# False SOLUTION: True, Batch normalization is typically applied to the inputs of a layer, normalizing them before they are passed through the activation function. 1.4 Dropout regularization can be applied only to fully connected layers in a neural network. \# True \# False SOLUTION: False, While the lecture slides specifically mention an example of dropout in a 3-layer network, they do not restrict its application to fully connected layers only. Dropout can be applied to different types of layers in a neural network 1.5 Saliency maps in convolutional neural networks (CNNs) are primarily used to reduce the dimensionality of the output layer for easier interpretation and visualization. \# True \# False SOLUTION: False, This question is relevant to the topic of saliency maps as discussed in the lecture, and the statement is false because saliency maps are actually used to highlight the areas of an input image that are most relevant to the network’s decision, rather than for reducing dimensionality of the output layer. 2 Multiple Choices (8 points) -Recommended 6 Minutes Fill in the circle next to the letter(s) of your choice (like this: ) . No explanations are required. Choose ALL options that apply. Each question is worth 4 points and the answer may contain one or more options. Selecting all of the correct options and none of the incorrect options will get full credits. For questions with multiple correct options, each incorrect or missing selection gets a 2-point deduction (up to 4 points) . 2.1 During the training of a deep neural network using Stochastic Gradient Descent (SGD), you notice that the model’s performance on the validation set starts declining after a certain number of epochs, eventhough the loss on the training set is still going down and the performance continues to improve. What would you do to improve the model’s training (Treat each option as independent of each other)? Select ALL options that apply. \# A: Decrease the regularization parameter 􀀕 . \# B: Reduce the number of layers in the neural network. \# C: Increase the number of neurons on each layer of the neural network. \# D: Increase the dropout rate. \# E:

FAQs

How is the quiz scored for True/False questions?

Correct answers are worth 2 points. Incorrect answers are worth -1 to discourage guessing, and leaving a question blank gives 0 points.

What is the core issue described in the ReLU dead-neuron statement?

When ReLU inputs are less than zero, outputs become zero, which can stop neurons from responding and lead to the dead ReLU problem.

Which multiple-choice actions are recommended to address overfitting in training?

The solutions select B (reduce the number of layers) and D (increase the dropout rate) as the applicable remedies.

Sep 30, 2026

Upload by supergirl