Convolutional neural networks

Question 1

Now, imagine that this 4×4×2 volume gets fed into a subsequent layer with 3 filters, each of shape 2×3×2. What will be the shape of the resulting output volume?

Answer 1

3×2×3

suppose you have an n₁×n₂×3 image and you f₁×f₂×3 and 3 filters with stride s, then the formula is (n₁-f₁)/s +1 × (n₂-f₂ )/s+1 × #amount of filters

Question 2

Describe the layers in convet

Answer 2

• Repeat following pattern
  
  • Conv layer
  • ReLu
  • Max pool
• Final layer is a Fully connected layer which determines the output (use softmax for multiclass classification)

Question 3

Describe max pooling

Answer 3

• Take the maximum within a region and only keep this element
• Pooling doesnt care where the maximum comes from, therefore makes it more robust to position

Question 4

Mention all of the CNN architectures

Answer 4

1. Lenet5
2. Alexnet
3. VGG
4. GoogleNet
5. ResNet

Question 5

Mention the novelty in each CNN architecture

Answer 5

Question 6

Describe the architecture behind AlexNet

Answer 6

Question 7

pixel space vs feature space

Give an example of images closei n pixel space and in feature space

Answer 7

• Pixel space is how close the RAW
  
  • i.e if pictures share large blue backgroun, then they may be close in pixel space, but not in feature space since the background does not matter for the object
• Feature space is how close the pictures according to the training labels
  
  • What do the pictures have in common in they all share same label
  • Abstracting away backround etc..