CNN

Question 1

What should be the padding to retain ‘same’ size of Convolution? Size is N and filter is F

Answer 1

P = (F-1)/2

Question 2

With NxN image, filter of size F, padding P, stride S, what is output size?

Answer 2

(N + 2P - F)/S + 1

Question 3

What is difference between Cross-Correlation and Convolution

Answer 3

When we apply filter in CNN, it is called ‘Cross Correlation’ in maths. Convolution is Math is Flipping the filter horizontally and vertically.

Question 4

If we have 10 filters with size=3 and channels=3, how many parameters in that layer

Answer 4

(333 + 1) * 10 = 280

Question 5

Why are CNNs less prone to over-fitting

Answer 5

Whatever be the size of image, the number of features can be quite limited (based on filter size and number of filters)

Question 6

What does max pooling achieve?

Answer 6

Preserves a feature found in a pool, and ignores the rest

Question 7

As we go deeper in a network, Height and Width _______ and number of layers________

Answer 7

Height and Width decrease and number of layers increase

Question 8

Lenet-5 Network: Describe size of each layer - (Conv->Pool->Conv->Pool->FC->FC), starting with 32X32X1

Answer 8

Check notes

Question 9

How many total parameters in Lenet-5?

Answer 9

60K

Question 10

What is a Resnet?

Answer 10

Activation from Layer l is added to Z(l +n) (before the activtion)

Question 11

In a CNN (vs Resnet), as the number of layers increase, the training error keeps on decreasing. True/False

Answer 11

False. It comes down and then goes up as number of layers increase. But in a Resnet, it keeps decreasing

Question 12

What is a ‘network in network’ convolution?

Answer 12

It is a 1X1 filter with n_c channels

Question 13

How do you shrink number of channels?

Answer 13

Using 1X1 filter with lesser channels

Question 14

What is a Bottleneck Layer in Inception?

Answer 14

It is the 1X1 layer used to reduce the computational cost