CNN

Question 1

CV problems

Answer 1

classifications, object detection, neural style transfer

Question 2

how to detect edges? (vertical, horizontal, dia)

Answer 2

use filter that contains 1s and 0s like
1 0 -1
1 0 -1
1 0 -1
to detect vertical edges
the values for pixels at the ver edge will be v large/small (lighter/darker color) –> diff from other pixels
similar for horizontal

Question 3

what is sobel or sehorr filter in edge detection?

Answer 3

1 0 -1
2 0 -2
1 0 -1
focus on the center pix
sehorr
3 0 -3
10 0 -10
3 0 -3

Question 4

what is padding

Answer 4

To prevent losing information at the edges of the image (convolution shrinks mattrix), add 0s around the orignal matrix to pad
p can be 0 1 2…

Question 5

what the result dimension when apply a fxf filter on nxn mat?

Answer 5

n-f+1xn-f+1

Question 6

type of padding?

Answer 6

valid: no padding
same: output size same as input size
calculated based on input and filter size (f is usually odd)

Question 7

what is strided conv?

Answer 7

moving filter s steps
s can be 1 2 3…

Question 8

formula to cal output dimension

Answer 8

(n+2p-f)/s+1 round down

Question 9

how to do cross correlation (deconvolution)

Answer 9

rotate the filter clockwise then flip it horizontally
then do inputxfilter

Question 10

what is the condtion of input and filter when the are more than 1 channel?

Answer 10

the number of channels of input and filter must be the same
the number of channels of the output will be number of filters used

Question 11

what is pooling layee? what are the types of pooling?

Answer 11

at pooling layer, instead of performing multiplication (*), use MAX, MIN, AVG operation instead.

Question 12

diff between conv layer and pool layer?

Answer 12

conv layer has params but pool layer doesnt
in NN, conv1+pool1–> layer 1, conv2+pool2–> layer 2…

Question 13

what is parameter sharing?

Answer 13

a feature detecter eg vertical edge detecter can be applied to other image to detect vertical edge

Question 14

why doing convolution?

Answer 14

parameter sharing
sparsity of connections

Question 15

what is sparsity of connections in convolution?

Answer 15

in each layer, each output value depends on small number of inputs

Question 16

what are some classic neural networks?

Answer 16

lenet5: conv1 avgpool1 conv2 avgpool2 fc1 fc2 softmax –> very simple, common type of arrangement

alexnet: bigger, use maxpool instead of avgpool, same arrangement

vgg16: use padding (same) to preserve output dim,

Question 17

what is residual block?

Answer 17

short cut from early input to later output (before ReLU sum up the ealier and the current)

Question 18

advantage of resnet

Answer 18

allow to train deeper NN without hurting generability

Question 19

why resnet work well without hurting performance?

Answer 19

identity function is ez to learn –> get back past result
g(wa+b +pasta)
because wa+b is small –> past a is large –> got a

Question 20

feature of resnet?

Answer 20

have skip connections

Question 21

what does an 1x1 convolution do?

Answer 21

shrink the number of channels or increase it

Question 22

motivation for inception network?

Answer 22

improve computational cost

Question 23

what is inception block?

Answer 23

apply different filters on 1 input and concat them as the output (bottle neck) –> input to other layer

Question 24

what is advantage of mobilenet?

Answer 24

no need large computational resources

Question 25

what is the feature of mobilenet?

Answer 25

depthwise separable conv
depthwise: filter will be fxf (not x nc) and nc filters ( 1 channel filter but no filters = no. channels) –> output has same nc
followed by pointwise conv with 1x1xnc’ filter
–> final output is n’xn’xnc’

Question 26

what is effnet?

Answer 26

width, height and resolution can be scaled uniformly (compound scaling)