Scribd
Upload
222 views2 pages

CISC 867: Deep Learning Assignment #2 (2 Points/question)

This document contains 13 questions about concepts in deep learning including convolutional neural networks, pooling layers, batch normalization, dropout, and training curves. The questions cover topics such as layer parameters, receptive field sizes, convolution operations, advantages of certain layers over fully connected layers, and how dropout works during training and testing.

Uploaded by

Yomna Salem
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
222 views2 pages

CISC 867: Deep Learning Assignment #2 (2 Points/question)

This document contains 13 questions about concepts in deep learning including convolutional neural networks, pooling layers, batch normalization, dropout, and training curves. The questions cover topics such as layer parameters, receptive field sizes, convolution operations, advantages of certain layers over fully connected layers, and how dropout works during training and testing.

Uploaded by

Yomna Salem
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd

CISC 867: Deep Learning

Assignment #2 (2 points/question)

1. Consider an input image of shape 500x500x3. The image is flattened and a fully connected layer
with 100 hidden units is used. What is the shape of the weight matrix of this layer (without the
bias)? What is the shape of the bias.

2. You run this image in a convolutional layer with 10 filters, of kernel size 5x5. How many arameters
does this layer have?

3. The top gray image has run through different types of filters and the results are shown in the
following images. What type of convolutional filter was used to get each of the resulting images.
Explain briefly and include the values of these filters. The filters have a shape of (3,3).

4. In the Adam optimizer. Show what will happen the numbers of steps to compute the exponential
moving averages gets large.

5. Given a batch of size m, and assume that a batch normalization layer takes an input z =
(𝑧 (1) , … , 𝑧 (𝑚) ) as an input. Write down the output equation(s) of this layer. Give two reasons for
using the batch normalization layer.

6. Suppose you have a convolutional network with the following architecture: The input is an RGB
image of size 256 x 256. The first layer is a convolution layer with 32 feature maps and filters of size
3x3. It uses a stride of 1, so it has the same width and height as the original image. _ The next layer
is a pooling layer with a stride of 2 (so it reduces the size of each dimension by a factor of 2) and
pooling groups of size 3 x 3.
Determine the size of the receptive _eld for a single unit in the pooling layer. (i.e., determine the
size of the region of the input image which influences the activation of that unit.) You may assume
the receptive field lies entirely within the image.

7. If an input data block in a convolutional network has dimension C×H×W = 96 ×128 ×128 , (96
channels, spatial dim 128x128) and we apply a convolutional filter to it of dimension D × C × HF
× WF = 128×96 ×7×7 , (i.e. a block of D=128 filters) with stride 2 and pad 3,what is the dimension
of the output data block?

8. What is inverted dropout and what is its advantage?

9. Explain briefly why fully connected neural networks do not work well for image classification.

10. Compute the convolution of the following two arrays: (4 1 -1 3) * (-2 1). Your answer should be an
array of length 5. Show your detailed work.

11. Describe what setting change in epochs 25 and 60 could have produced this training curve. Be brief.

12. Why are convolutional layers more commonly used than fully-connected layers for image
processing?

13. Dropout layers implement different forward functions at train and test time. Explain what they do.
Let p be the probability that node value is retained.

You might also like