C4_W2.pdf

Case Studies

Classic Networks (architectures)

LeNet-5

AS you go deeper in the network the height and width of the images tend to go down but the number of channels tend to go up

Conv ⇒ Avg Pool ⇒ Conv ⇒ Avg Pool ⇒ FC ⇒ FC ⇒ Output

60k parameters

Uses Sigmoid/TanH activation function

AlexNet (Easiest paper to read)

Conv ⇒ Max Pool ⇒ 5x5 Same Conv ⇒ Max Pool ⇒ 3x3 Same Conv ⇒ 3x3 Same Conv ⇒ ⇒ 3x3 Same Conv ⇒ Max Pool ⇒ FC ⇒ FC ⇒ Softmax

60M parameters

Uses RELU activation function

Local Response Normalization

VGG - 16

138M parameters

The architecture is quite uniform and simple