C4W2

Question 1

Some important ConvNets

Answer 1

LeNet-5(60k)
AlexNet(60 millions)
VGG(138 millions)

Question 2

What is ResNet?

Answer 2

Network that contains Residual Block

Question 3

What is Residual Block?

Answer 3

ResBlock or Skip Connection or Shortcut - adds output of one layer to another skipping the layer in between.
It doesn’t hurt performance and helps to train more deep NN

Question 4

What to use 1x1 filter?

Answer 4

When you want to adjust number of channels

Question 5

What is inception network?

Answer 5

When you apply different filters and layers (Conv/Pool) and then stack their output

Question 6

What is MobileNet?

Answer 6

NN optimised for constrained environment such as Mobile Phone

Question 7

What is Deep wise separable convolution?

Answer 7

When you apply each filter to only 1 channel

Question 8

What is pointwise convolution?

Answer 8

Use 1x1 filter to output desired number of channels

Question 9

What is the idea of MobileNet v2? What is expansion layer?

Answer 9

MobileNetV2 uses residual blocks and expansion layers.
Expansion layer expands the input by several 1x1 filters, then applies deepwise separable conv, then pointwise conv

Question 10

What is EfficientNet?

Answer 10

NN architecture which scales:
1. Resolution
2. NN depth
3. NN width

Question 11

How to get good performance when you have lack of data for image classification task?

Answer 11

Use open source implementation and weight
Retrain the last SoftMax layer and optionally couple of the latest layers
Use data augmentation, use color shifting

Question 12

How to do good on benchmarks?

Answer 12

Ensembling - average output of several NN
Multicrop on test time - run NN on multiple versions of the test image