Description
Note: All coding problems to be submited with Github Link. Do not Upload the files/folder. Use git commands only.
Note: this is the distribution of questions:
(a) Question 1 to Question2: Required for everyone.
(b) Question 3: Required only for Graduate Students
(c) Question 4: Bonus question for both Graduate Students and Undergraduate Students
Problem 1 (20 points)
You have a convolutional neural network that takes as an input image of size 512 × 512 × 3 and passes it through a layer that convolves the image using 3 filters of dimensions 5 × 5 × 3 with a valid padding.
(a) List all learnable parameters of this convolution layer.
(b) What if you want to replicate the behavior of this convolutional layer using a fully connected layer? How many parameters would that fully connected layer have?
Problem 2 (20 points)
Given a binary input image of diagonal streaks (see example in Figure 2) and two filters (see Figure 1a) describe how would you build a detector for finding the location of pattern shown in Figure 1b on the input image. Allowed operations are convolution, summation, and argmax.
Bonus for undergraduates beyond this line
Problem 3 (20 points)
Demonstrate that convolution is translation invariant for 1D convolution (Note:this can be extended to N-D convolutions as well).
Bonus for both undergraduates and gradu-
ates beyond this line.
1
0
1
2
3
4
5
6
0 2 4 6
0
0
1
1
2
2
3
3
4
4
5
5
6
6
0
2
4
6
0
1
2
3
4
5
6
(a) Available fixed filters (b) Target shape to detect
Figure 1: Filters to use (1a) and pattern to find (1b).
Problem 4 (20 points)
You have to choose between two papers given below:
(a) Paper 1: High-Performance Neural Networks for Visual Object Classifica-tion:
(i) Give a short summary of the paper.
(ii) What were the parameter sizes for CIFAR-10 and MNIST? why do you think the paramtere size differed for CIFAR-10 vs MNIST?
(b) Paper 2: ImageNet Classification with Deep Convolutional Neural Networks:
(i) Give a short summary of the paper.
(ii) Why is there a big fluctuation of loss for the last epoch of training?
2
0
100
200
300
400
500
600
700
0 100 200 300 400 500 600 700
Figure 2: An example image for the architecture
3
