Convonet

Question 1

What is the curse of dimensionality?

Answer 1

Objects of the same “size” will make up less and less of the space as the dimension increase. This can be visualised as a ball inscribed in a cube makes up less and less of the volume as the dimension increases.

Question 2

What is the problem of self-similarity?

Answer 2

Different areas in a image might seem the same when viewed out of context.

Question 3

What are the main differences of the Lenet and the Alex net?

Answer 3

Lenet:
Tanh
Average pooling
Small filters (5x5

Alexnet:
Relu
Max pooling
Data augmentation
Large filters (11x11)

Question 4

What is the ouput size of a concolution layer?

Answer 4

1 + (in_size + 2*padding - filter_size) / stride

Question 5

What is the main difference from AlexNet to VGG?

Answer 5

Smaller filters (3x3)
Deeper (16 layers)
Uniform architecture (All layers the same).

Question 6

What is the main feature of GoogLeNet?

Answer 6

Inception blocks, several different convolutions in each block. The network can “learn” which block to use. In addition it uses auxilary training, very small filter sizes , batch normalization.

Question 7

What is the main feature of ResNet?

Answer 7

Skip connections, allows gradient propagation trough extremly deep networks.