CSE 493 G1/599 G1 Deep Learning Spring 2025 Practice Exam - SOLUTIONS

Sep 30, 2026

Upload by McQueen

Deep Learning practice exam solutions for CSE 493 G1 and 599 G1 (Spring 2025), dated May 15, 2025. The document presents a structured answer key for multiple question types including True/False items and a multi-choice section, with scoring notes and instructions about how to mark responses. Solutions explain key concepts such as feature learning from raw pixels, behavior of Softmax loss, learning-rate effects, convexity under SGD, vanishing gradients, convolutional expressivity, and relationships between test loss and accuracy.

LanguageEnglish

CategoryExam

FormatPDF

Pages17

Reading Time43 min

AccessFree to download

True/False

Question 1.1 - 1.10

Multiple Choices

Question 2.1

CSE 493 G1/ 599 G1 Deep Learning Spring 2025 Practice Exam SOLUTIONS May 15, 2025 Full Name: UW Net ID: | Question | Score | | --- | --- | | True/False (20 pts) | | | --- | --- | | Multiple Choice (40 pts) | | | Backpropagation (20 pts) | | | Convolution & Pooling (20 pts) | | | Total (100 pts) | | Welcome to the CSE 493 G1 Exam! • The exam is 80 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 (20 points) 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 Training on higher-level image features (e.g. color histograms or Histogram of Gradient features) instead of raw pixels improves image classification accuracy across all types of models. \# True \# False SOLUTION: False. While we saw better performance in A1 for an SVM and a two-layer neural network, with more expressive models like neural networks (and sufficient training data), it is often better to allow the model to learn its own features from raw pixels directly. 1.2 If your datapoints are linearly separable, then training a linear classifier with a Softmax loss and no regularization will achieve zero training loss at convergence. \# True \# False SOLUTION: False. The output of the softmax function only achieves 1 in the limit as values to go +infinity, and therefore in practice a loss of exactly 0 will not be observed. 1.3 During training, you examine the loss curve, and you notice that your training loss is plateauing out very early at a high value. One valid method to try is to increase the learning rate. \# True \# False SOLUTION: FALSE: The learning curve indicates that the learning rate is already too high, so increasing it would not help. 1.4 If the loss function is convex, Stochastic Gradient Descent always reaches the global minimum. \# True \# False SOLUTION: False, when step size is large, it is not guaranteed 1.5 Unlike the sigmoid activation function, tanh does not suffer from vanishing gradients. \# True \# False SOLUTION: False. The backpropagated gradient for tanh is in fact less than or equal to the upstream gradient so it can cause gradients to die in deeper networks. 1.6 A convolutional layer can express any function a fully connected layer can express. \# True \# False SOLUTION: False. It’s the opposite: fully connected layers are actually strictly more expressive than convolutional layers, which are a particular instance of a weight matrix multiplication. For example, a fully connected layer could learn to process the top pixels of an image differently than the bottom pixels, while a convolutional layer applies the same kernel to all regions of the image. This is usually a good thing, since we want to detect an object the same way regardless of where it is in the image. It is this property of convolutions that makes them efficient and generalizable, but we are necessarily giving up some expressivity. Note that this question is asking about the general case, so you cannot assume for example that the kernel is the same size as the input. 1.7 If Model A has a lower test loss on a dataset than Model B, then Model A must have a higher accuracy on the test dataset than Model B. \# True \# False SOLUTION: False. Consider a dataset of 2 images, image 1 has ground truth value of Cat and and Image 2 has ground truth value of Dog. Let Model A and Model B have SVM loss with margin 1 . Model A says that the first image score is [Dog: 1.1, Cat: 1.0] and the second image score is [Dog: 1.0, Cat: 1.1] . I

FAQs

What are the scoring rules for the True/False section?

Each True/False question is worth 2 points for a correct answer. Incorrect answers are penalized with -1 points, and leaving a question blank gives 0 points.

Why is the statement about Softmax loss achieving exactly zero training loss considered false?

The Softmax output reaches 1 only in the limit as scores approach +infinity, so in practice a loss of exactly 0 is not observed.

Does higher test loss necessarily imply lower accuracy on the test set?

No. The solutions explain that different models can have different loss values without a direct one-to-one mapping to accuracy, using a concrete example with SVM margins and predicted scores.

Sep 30, 2026

Upload by McQueen