Encoder Decoder Networks

Question 1

What tasks can be accomplished by decoding

Answer 1

Classofication
mapping
captioning
translation
generation
recognition

Question 2

Define decoding

Answer 2

Process of decoding the information represented into a vector whose elements are extracted from predefined alphabet

Question 3

For what can interpolation maps be used?

Answer 3

Up-sampling from a low-resolution image towards an high resolution image

Question 4

what is the deep learning optimal approach to up-sampling and why?

Answer 4

Transposed convolution
It does not use a predefined interpolation method. Learning will allow the tuning of the parameters

Question 5

Transposed convolution is a many-to-one association T or F

Answer 5

F

Question 6

Transposed convolution has the same connectivity as the normal convolution in the backward direction T or F

Answer 6

T

Question 7

Transposed convolution is a type of convolution T or F

Answer 7

F

Question 8

How can we emulate transposed convolution using convolution?

Answer 8

By up-sampling the input by adding zeros between the values in the input matrix in a way that the direct convolution produces the same effect as the transposed convolution

Question 9

Weights in the transposed convolution are learnable T or F

Answer 9

T

Question 10

For up-sampling with transposed convolution we need a predefined interpolation method T or F

Answer 10

F

Question 11

Other names for transposed convolution?

Answer 11

Deconvolution
Fractional strided convolution

Question 12

Why transposed convolution is called so?

Answer 12

Because if we interpret convolution as a matrix product between a rearranged version of the filter and the image, than TC correspond to a matrix multiplication between the transposed matrix of the rearranged filter and the image to up-sample

Question 13

How to perform transposed convolution

Answer 13

Match the filter with a single pixel of the image by scalar multiplication. Than repeat for the next pixel, the position of the result in the output is according to the strided. Perform sum on the superposition of the output

Question 14

Possible output layers for encoder-decoder networks

Answer 14

Linear, binary, softmax

Question 15

Main feature of U-Net

Answer 15

Skipe connections between correspondent encoding and decoding layer. The outpur of progressive decoding step is concatenated to the correspondent feature map in the encoding layer and the receptive field fo the neuron spans the full depth of the result of the concatenation

Question 16

sequence of layers in an encoding block of the U-Net

Answer 16

Conv, Relu, MaxPool

Question 17

sequence of layers in decoding block of U-Net

Answer 17

Deconv, Relu, Concat, Conv, Relu

Question 18

Segmentation can be framed as a classification task T or F

Answer 18

T

Question 19

In U-Net for segmentation, what is the dimension of the output

Answer 19

The dimensions of the input plus an extra dimension equal to the number of classes

Question 20

In U-Net for segmentation, what is encoded in the first feature maps?

Answer 20

Sharpening, Edge and boundaries, Emboss/Deboss

Question 21

U-Net for segmentation, what is encoded in the deeper feature maps?

Answer 21

Shape clustering

Question 22

U-Net for segmentation, what is decoded in the first layes?

Answer 22

On-Off semantics

Question 23

U-Net for segmentation what is performed deeper layer of decoding

Answer 23

Background neutralization, adversarial mapping

Question 24

Where is the true intelligence of encoder-decoder network?

Answer 24

In the decoder network