You are on page 1of 492

CS7015 (Deep Learning) : Lecture 12

Visualizing Convolutional Neural Networks, Guided Backpropagation, Deep


Dream, Deep Art, Fooling Convolutional Neural Networks

Mitesh M. Khapra

Department of Computer Science and Engineering


Indian Institute of Technology Madras

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 1/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Acknowledgements
Andrej Karpathy Video Lecture on Visualization and Deep Dream∗


Visualization, Deep Dream, Neural Style, Adversarial Examples

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 2/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Module : Visualizing patches which maximally activate
a neuron

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 3/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
55

27
23
11
11 9
3 5 7
5 dense
3 3 3 3 3
3 3 2 dense dense
3 3 3
5 3
11 5 2
7
256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution
384
256
Convolution
MaxPooling 1000
256
96
Convolution
MaxPooling
4096 4096
96

Convolution

Input

Mitesh M. Khapra
. .
.
layer of a CNN

CS7015 (Deep Learning) : Lecture 12


. . . . . . . . . . . . . .
. . . . . . . . . . . . . .
.
. .
. . .
. . .
Consider some neurons in a given

4/1
1000
dense
Consider some neurons in a given

4096
dense
layer of a CNN

4096
dense

MaxPooling
We can feed in images to this CNN

256
5

Convolution
and identify the images which cause

256
7

7
these neurons to fire

Convolution
384
9

Convolution
384
11

11

MaxPooling
23

256
23

Convolution
27

256
27
55

MaxPooling
96
55

Convolution
96

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 4/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
1000
dense
Consider some neurons in a given

4096
dense
layer of a CNN

4096
dense

MaxPooling
We can feed in images to this CNN

256
5

Convolution
and identify the images which cause

256
7

7
these neurons to fire

Convolution
384
9

9
We can then trace back to the patch

Convolution
384
11
in the image which causes these neur-
11

MaxPooling
23

256
ons to fire
23

Convolution
27

256
27
55

MaxPooling
96
55

Convolution
96

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 4/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
1000
dense
Consider some neurons in a given

4096
dense
layer of a CNN

4096
dense

MaxPooling
We can feed in images to this CNN

256
5

Convolution
and identify the images which cause

256
7

7
these neurons to fire

Convolution
384
9

9
We can then trace back to the patch

Convolution
384
11
in the image which causes these neur-
11

MaxPooling
23

256
ons to fire
23

Convolution
27

256
27
55

MaxPooling
96
55

Convolution
96

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 4/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
1000
dense
Consider some neurons in a given

4096
dense
layer of a CNN

4096
dense

MaxPooling
We can feed in images to this CNN

256
5

Convolution
and identify the images which cause

256
7

7
these neurons to fire

Convolution
384
9

9
We can then trace back to the patch

Convolution
384
11
in the image which causes these neur-
11

MaxPooling
23

256
ons to fire
23

Convolution
27

256
27
55

MaxPooling
96
55

Convolution
96

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 4/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
1000
dense
Consider some neurons in a given

4096
dense
layer of a CNN

4096
dense

MaxPooling
We can feed in images to this CNN

256
5

Convolution
and identify the images which cause

256
7

7
these neurons to fire

Convolution
384
9

9
We can then trace back to the patch

Convolution
384
11
in the image which causes these neur-
11

MaxPooling
23

256
ons to fire
23

Convolution
Let us look at the result of one of such
27

256
27

experiment conducted by Grishick et


55

MaxPooling
96
55

Convolution

al., 2014
96

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 4/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
They consider 6 neurons in the pool5 layer and find the image patches which
cause these neurons to fire
One neuron fires for people faces

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 5/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
They consider 6 neurons in the pool5 layer and find the image patches which
cause these neurons to fire
Another neuron fires for dog faces

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 5/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
They consider 6 neurons in the pool5 layer and find the image patches which
cause these neurons to fire
Another neuron fires for flowers

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 5/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
They consider 6 neurons in the pool5 layer and find the image patches which
cause these neurons to fire
Another neuron fires for numbers

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 5/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
They consider 6 neurons in the pool5 layer and find the image patches which
cause these neurons to fire
Another neuron fires for houses

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 5/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
They consider 6 neurons in the pool5 layer and find the image patches which
cause these neurons to fire
Another neuron fires for shiny surfaces

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 5/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Module : Visualizing filters of a CNN

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 6/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Recall that we had done something
x̂ similar while discussing autoencoders

h(x)

max {wT x}
x

s.t. ||x||2 = xT x = 1
w1
Solution: x = √
w1T w1

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 7/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Recall that we had done something
x̂ similar while discussing autoencoders
We are interested in finding an input
which maximally excites a neuron
h(x)

max {wT x}
x

s.t. ||x||2 = xT x = 1
w1
Solution: x = √
w1T w1

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 7/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Recall that we had done something
x̂ similar while discussing autoencoders
We are interested in finding an input
which maximally excites a neuron
h(x)
Turns out that the input which will
W
maximally activate a neuron is ∥W ∥

max {wT x}
x

s.t. ||x||2 = xT x = 1
w1
Solution: x = √
w1T w1

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 7/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
h11 h12 Now recall that we can think of a
CNN also as a feed-forward network
with sparse connections and weight
sharing
...
16

2
h14
* =

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 8/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
h11 h12 Now recall that we can think of a
CNN also as a feed-forward network
with sparse connections and weight
sharing
...
Once again, we are interested in
16 knowing what kind of inputs will
cause a given neuron to fire

2
h14
* =

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 8/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
h11 h12 Now recall that we can think of a
CNN also as a feed-forward network
with sparse connections and weight
sharing
...
Once again, we are interested in
16 knowing what kind of inputs will
cause a given neuron to fire

The solution would be the same

2
h14
∥ ) where W is the filter(2 × 2, in
W
( ∥W
* =
this case)

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 8/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
h11 h12 Now recall that we can think of a
CNN also as a feed-forward network
with sparse connections and weight
sharing
...
Once again, we are interested in
16 knowing what kind of inputs will
cause a given neuron to fire

The solution would be the same

2
h14
∥ ) where W is the filter(2 × 2, in
W
( ∥W
* =
this case)

We can thus think of these filters as


pattern detectors
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 8/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We can simply plot the K ×K weights
(filters) as images & visualize them as
patterns

max {wT x}
x

s.t. ||x||2 = xT x = 1
w1
Solution: x = √
w1T w1

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 9/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We can simply plot the K ×K weights
(filters) as images & visualize them as
patterns
The filters essentially detect these
patterns (by causing the neurons to
maximally fire)

max {wT x}
x

s.t. ||x||2 = xT x = 1
w1
Solution: x = √
w1T w1

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 9/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We can simply plot the K ×K weights
(filters) as images & visualize them as
patterns
The filters essentially detect these
patterns (by causing the neurons to
maximally fire)
This is only interpretable for the fil-
ters in the first convolution layer
max {wT x}
x

s.t. ||x||2 = xT x = 1
w1
Solution: x = √
w1T w1

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 9/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We can simply plot the K ×K weights
(filters) as images & visualize them as
patterns
The filters essentially detect these
patterns (by causing the neurons to
maximally fire)
This is only interpretable for the fil-
ters in the first convolution layer
max {wT x} (Why?)
x

s.t. ||x||2 = xT x = 1
w1
Solution: x = √
w1T w1

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 9/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Module : Occlusion experiments

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 10/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
pomeranian wheel ... hound
Typically we are interested in under-
Softmax
standing which portions of the image
(a) Input Image (b) Layer 5, strongest feature map
(c) Layer 5, strongest
feature map projections
are responsible for maximizing the
(d) Classifier, probability
of correct class
(e) Classifier, most
probable class

probability of a certain class


0.9

0.8
Pomeranian
Tennis ball
Keeshond
Pekinese

0.7

0.6

0.5

0.4

0.3

0.2

True Label: Pomeranian 0.1

Car wheel
Racer
0.25 Cab
Police van

0.2

0.15

0.1

0.05

True Label: Car Wheel


Afghan hound
0.7
Gordon setter
Irish setter
0.6 Mortarboard
Fur coat
0.5
Academic gown
Australian terrier
Ice lolly
0.4
Vizsla
Neck brace
0.3

0.2

0.1
True Label: Afghan Hound

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 11/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
pomeranian wheel ... hound
Typically we are interested in under-
Softmax
standing which portions of the image
(a) Input Image (b) Layer 5, strongest feature map
(c) Layer 5, strongest
feature map projections
(d) Classifier, probability
of correct class
are responsible for maximizing the
(e) Classifier, most
probable class

probability of a certain class


0.9

0.8
Pomeranian
Tennis ball
Keeshond
Pekinese

0.7

We could occlude (gray out) differ-


0.6

0.5

0.4

True Label: Pomeranian


ent patches in the image and see the
0.3

0.2

0.1

effect on the predicted probability of


0.25
Car wheel
Racer
Cab
Police van

the correct class


0.2

0.15

0.1

0.05

True Label: Car Wheel


Afghan hound
0.7
Gordon setter
Irish setter
0.6 Mortarboard
Fur coat
0.5
Academic gown
Australian terrier
Ice lolly
0.4
Vizsla
Neck brace
0.3

0.2

0.1
True Label: Afghan Hound

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 11/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
pomeranian wheel ... hound
Typically we are interested in under-
Softmax
standing which portions of the image
(a) Input Image (b) Layer 5, strongest feature map
(c) Layer 5, strongest
feature map projections
(d) Classifier, probability
of correct class
are responsible for maximizing the
(e) Classifier, most
probable class

probability of a certain class


0.9

0.8
Pomeranian
Tennis ball
Keeshond
Pekinese

0.7

We could occlude (gray out) differ-


0.6

0.5

0.4

True Label: Pomeranian


ent patches in the image and see the
0.3

0.2

0.1

effect on the predicted probability of


0.25
Car wheel
Racer
Cab
Police van

the correct class


0.2

0.15

For example this heat map shows that


0.1

0.05

True Label: Car Wheel


occluding the face of the dog causes
0.7
Afghan hound
Gordon setter
Irish setter
0.6 Mortarboard

a maximum drop in the prediction


0.5

0.4
Fur coat
Academic gown
Australian terrier
Ice lolly
Vizsla
Neck brace

probability
0.3

0.2

0.1
True Label: Afghan Hound

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 11/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
pomeranian wheel ... hound
Typically we are interested in under-
Softmax
standing which portions of the image
(a) Input Image (b) Layer 5, strongest feature map
(c) Layer 5, strongest
feature map projections
(d) Classifier, probability
of correct class
are responsible for maximizing the
(e) Classifier, most
probable class

probability of a certain class


0.9

0.8
Pomeranian
Tennis ball
Keeshond
Pekinese

0.7

We could occlude (gray out) differ-


0.6

0.5

0.4

True Label: Pomeranian


ent patches in the image and see the
0.3

0.2

0.1

effect on the predicted probability of


0.25
Car wheel
Racer
Cab
Police van

the correct class


0.2

0.15

For example this heat map shows that


0.1

0.05

True Label: Car Wheel


occluding the face of the dog causes
0.7
Afghan hound
Gordon setter
Irish setter
0.6 Mortarboard

a maximum drop in the prediction


0.5

0.4
Fur coat
Academic gown
Australian terrier
Ice lolly
Vizsla
Neck brace

probability
0.3

0.2

0.1
True Label: Afghan Hound
Similar observations are made for
other images

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 11/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Module : Finding influence of input pixels using
backpropagation

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 12/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We can think of an image as a m × n
inputs x0 , x1 , . . . , xm×n

flatten

x0 x1 · · · ·
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · xmn
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 13/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We can think of an image as a m × n
inputs x0 , x1 , . . . , xm×n
hj
We are interested in finding the influ-
ence of each of these inputs(xi ) on a
given neuron(hj )

xi

flatten

x0 x1 · · · ·
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · xmn
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 13/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We can think of an image as a m × n
inputs x0 , x1 , . . . , xm×n
hj
We are interested in finding the influ-
ence of each of these inputs(xi ) on a
given neuron(hj )
If a small change in xi causes a large
change in hj then we can say that xi
xi
has a lot of influence of hj
flatten

x0 x1 · · · ·
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · xmn
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 13/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We can think of an image as a m × n
inputs x0 , x1 , . . . , xm×n
hj
We are interested in finding the influ-
ence of each of these inputs(xi ) on a
given neuron(hj )
If a small change in xi causes a large
change in hj then we can say that xi
xi
has a lot of influence of hj
flatten ∂hj
In other words the gradient ∂xi could
x0 x1 · · · · tell us about the influence
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · xmn
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 13/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
∂hj
=0 −→ no influence
hj ∂xi

xi

flatten

∂hi ∂hi
∂x0 ∂x1 · · · ·
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · ∂x∂hmn
i

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 14/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
∂hj
=0 −→ no influence
hj ∂xi
∂hj
= large −→ high influence
∂xi

xi

flatten

∂hi ∂hi
∂x0 ∂x1 · · · ·
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · ∂x∂hmn
i

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 14/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
∂hj
=0 −→ no influence
hj ∂xi
∂hj
= large −→ high influence
∂xi
∂hj
= small −→ low influence
∂xi
xi

flatten

∂hi ∂hi
∂x0 ∂x1 · · · ·
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · ∂x∂hmn
i

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 14/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
∂hj
=0 −→ no influence
hj ∂xi
∂hj
= large −→ high influence
∂xi
∂hj
= small −→ low influence
∂xi
xi

flatten We could just compute these partial


derivatives w.r.t all the inputs
∂hi ∂hi
∂x0 ∂x1 · · · ·
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · ∂x∂hmn
i

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 14/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
∂hj
=0 −→ no influence
hj ∂xi
∂hj
= large −→ high influence
∂xi
∂hj
= small −→ low influence
∂xi
xi

flatten We could just compute these partial


derivatives w.r.t all the inputs
∂hi ∂hi
∂x0 ∂x1 · · · · And then visualize this gradient mat-
· · · · · · rix as an image itself
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · ∂x∂hmn
i

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 14/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
But how do we compute these gradi-
ents?

flatten

x0 x1 · · · ·
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · xmn
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 15/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
But how do we compute these gradi-
ents?
h32
Recall that we can represent CNNs
by feedforward neural network
h1i

x2

flatten

x0 x1 · · · ·
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · xmn
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 15/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
But how do we compute these gradi-
ents?
h32
Recall that we can represent CNNs
by feedforward neural network
h1i
Then we already know how to com-
pute influences (gradient) using back-
propogation
x2

flatten

x0 x1 · · · ·
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · xmn
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 15/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
But how do we compute these gradi-
ents?
h32
Recall that we can represent CNNs
by feedforward neural network
h1i
Then we already know how to com-
pute influences (gradient) using back-
propogation
x2
For example, we know how to back-
flatten prop the gradients till the first hidden
layer
x0 x1 · · · ·
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · ·
· · · · · xmn
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 15/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
But how do we compute these gradi-
ents?
h32
Recall that we can represent CNNs
by feedforward neural network
h1i
Then we already know how to com-
pute influences (gradient) using back-
propogation
x2
For example, we know how to back-
flatten prop the gradients till the first hidden
layer
x0 x1 · · · · ∑ ∂h32 ∂h1i3
∂h32
· · · · · · ∂x2
=
∂h1i ∂x2
· · · · · · i=1


4
· · · · · · h1i = wji xj
· · · · · · j=1

· · · · · xmn ∂h1i
= w12
∂x
.. 2
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 15/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
This is what we get if we compute the
gradients and plot it as an image

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 16/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
This is what we get if we compute the
gradients and plot it as an image
The above procedure does not show
very sharp influences

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 16/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
This is what we get if we compute the
gradients and plot it as an image
The above procedure does not show
very sharp influences
Springenberg et al. proposed “guided
back propagation” which gives a bet-
ter idea about the influences

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 16/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Module : Guided Backpropagation

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 17/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We feed an input to the CNN and do
a forward pass

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 18/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We feed an input to the CNN and do
a forward pass
We consider one neuron in some fea-
ture map at some layer

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 18/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We feed an input to the CNN and do
a forward pass
We consider one neuron in some fea-
ture map at some layer
We are interested in finding the influ-
ence of the input on this neuron

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 18/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We feed an input to the CNN and do
a forward pass
We consider one neuron in some fea-
ture map at some layer
We are interested in finding the influ-
ence of the input on this neuron
We retain this neuron and set all
other neurons in the layer to zero

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 18/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We now backpropogate all the way to
the inputs

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 19/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We now backpropogate all the way to
the inputs
1 -1 5 1 0 5 Recall that during forward pass relu
Forward pass
2 -5 -7 2 0 0 activation allows only positive values
-3 2 4 0 2 4 to pass & clamps −ve values to zero
-2 0 -1 -2 3 -1
Backward pass:
6 0 0 6 -3 1
backpropagation
0 -1 3 2 -1 3

0 3 0 -2 3 -1
Backward pass:
6 0 1 6 -3 1
“deconvnet”
2 0 3 2 -1 3

0 0 0 -2 3 -1
Backward pass:
guided 6 0 0 6 -3 1
backpropagation 0 0 3 2 -1 3 . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 19/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We now backpropogate all the way to
the inputs
1 -1 5 1 0 5 Recall that during forward pass relu
Forward pass
2
1 -5
-1 -7
5 2
1 0
0 0
5
activation allows only positive values
Forward pass -3
2 2
-5 4
-7 02 20 40 to pass & clamps −ve values to zero
-3 2 4 0 2 4
-2 0 -1 -2 3 -1 Similarly during backward pass no
Backward pass:
backpropagation
6
-2 0
0 0
-1 6
-2 -3
3 1
-1 gradient passes through the dead relu
Backward pass: 0 -1 30 26 -1 31
backpropagation
6 0 -3 neurons
0 -1 3 2 -1 3
0 3 0 -2 3 -1
Backward pass:
6
0 0
3 1
0 6
-2 -3
3 1
-1
“deconvnet”
Backward pass: 26 00 31 26 -1 31
-3
“deconvnet”
2 0 3 2 -1 3
0 0 0 -2 3 -1
Backward pass:
guided 6
0 0
0 0
0 6
-2 -3
3 1
-1
Backward pass:
backpropagation . . . . . . . . . . . . . . . . . . . .
06 00 30 26 -1 31
guided -3 . . . . . . . . . . . . . . . . . . . . 19/1
backpropagation 0 0 3 2 -1Mitesh
3 M. Khapra CS7015 (Deep Learning) : Lecture 12
We now backpropogate all the way to
1 -1 5 1 0 5
the inputs
Forward pass
2
1 -5
-1 -7
5 2
1 0
0 0
5 Recall that during forward pass relu
Forward pass -3
2
1
2
-5
-1
4
-7
5
02
1
20
0
40
5
activation allows only positive values
Forward pass -3
2
-2
2
-5
0
4
-7
-1
02
-2
20
3
40
-1
to pass & clamps −ve values to zero
Backward pass: -3 20 40 0 2 4
backpropagation
6
-2 0 -1 6
-2 -3
3 1
-1 Similarly during backward pass no
Backward pass: 0 -1 30 26 -1 31
backpropagation
6
-2 0
0 -1 -2 -3
3 -1 gradient passes through the dead relu
Backward pass: 0 -1 30 26 -1 31
backpropagation
6
0 0
3 0 -2 -3
3 -1 neurons
Backward pass: 06 -1 31 26 -1 31
0 -3
“deconvnet”
Backward pass:
0 3 0 -2 3 -1 In guided back propagation any -
26 00 31 26 -1
-3 31
“deconvnet” 0 3 0 -2 3 -1 ve gradients flowing from the upper
Backward pass: 26 00 31 26 -1 31
-3
“deconvnet”
Backward pass: 0 0 0 -2 3 -1 layer are also set to 0
guided 2
6 0
0 3
0 62 -1
-3 13
0 0 0 -2 3 -1
Backward pass:
backpropagation 06 00 30 26 -1 31
guided 0 0 0 -2 -3
3 -1
Backward pass:
backpropagation . . . . . . . . . . . . . . . . . . . .
06 00 30 26 -1 31
guided -3 . . . . . . . . . . . . . . . . . . . . 19/1
backpropagation 0 0 3 2 -1Mitesh
3 M. Khapra CS7015 (Deep Learning) : Lecture 12
Intuition: Neglect all the negative
Backpropagation influences (gradients) and focus only
on the positive influences (gradients)

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 20/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Intuition: Neglect all the negative
Backpropagation influences (gradients) and focus only
on the positive influences (gradients)
This gives a better picture of the true
influence of the input

Guided Backpropagation

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 20/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Module : Optimization over images

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 21/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Suppose we want to create an image
which looks like a dumbell (or an os-
55

27
trich, or a car, or just anything)
23
11 Dumbell
11 9
3 5 7
5 dense
3 3 3 3 3
3 3 2 dense dense
3 3 3
5 3
11 5 2
7
256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution
384
256
Convolution
MaxPooling 1000
256
96
Convolution
MaxPooling
4096 4096
96

Convolution

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 22/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Suppose we want to create an image
which looks like a dumbell (or an os-
55

27
trich, or a car, or just anything)
23
11 Dumbell
11
3

3
5
3
3
3
3
3
3
9

3
3
7

3
5 dense

2 dense dense
In other words we want to create
5 3

an image such that if we pass it


11 5 2
7
256
11 9 256
MaxPooling
27 23 384 Convolution

55
256

MaxPooling
384

Convolution
Convolution

1000
through a trained ConvNet it should
256
96

96
MaxPooling
Convolution
4096 4096 maximize the probability of the class
Convolution

Input
dumbell

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 22/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Suppose we want to create an image
which looks like a dumbell (or an os-
55

27
trich, or a car, or just anything)
23
11 Dumbell
11
3

3
5
3
3
3
3
3
3
9

3
3
7

3
5 dense

2 dense dense
In other words we want to create
5 3

an image such that if we pass it


11 5 2
7
256
11 9 256
MaxPooling
27 23 384 Convolution

55
256

MaxPooling
384

Convolution
Convolution

1000
through a trained ConvNet it should
256
96

96
MaxPooling
Convolution
4096 4096 maximize the probability of the class
Convolution

Input
dumbell
We could pose this as an optimization
problem w.r.t I (i0 , i1 , . . . , imn )

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 22/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Suppose we want to create an image
which looks like a dumbell (or an os-
55

27
trich, or a car, or just anything)
23
11 Dumbell
11
3

3
5
3
3
3
3
3
3
9

3
3
7

3
5 dense

2 dense dense
In other words we want to create
5 3

an image such that if we pass it


11 5 2
7
256
11 9 256
MaxPooling
27 23 384 Convolution

55
256

MaxPooling
384

Convolution
Convolution

1000
through a trained ConvNet it should
256
96

96
MaxPooling
Convolution
4096 4096 maximize the probability of the class
Convolution

Input
dumbell
We could pose this as an optimization
problem w.r.t I (i0 , i1 , . . . , imn )

arg max(Sc (I) − λΩ(I))


I
Sc (I) = Score for class C before softmax
Ω(I) = Some regularizer to ensure that
I looks like an image
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 22/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We can essentially think of the image
as a collection of parameters
55

27
23
11 Dumbell
11 9
3 5 7
5 dense
3 3 3 3 3
3 3 2 dense dense
3 3 3
5 3
11 5 2
7
256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution
384
256
Convolution
MaxPooling 1000
256
96
Convolution
MaxPooling
4096 4096
96

Convolution

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 23/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We can essentially think of the image
as a collection of parameters
55

27
23
Keep the weights of trained convolu-
11 Dumbell
11
3

3
5
3
3
3
3
3
3
9

3
3
7

3
3
5 dense

2 dense dense
tional neural network fixed
5
11 5 2
7
256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution
384
256
Convolution
MaxPooling 1000
256
96
Convolution
MaxPooling
4096 4096
96

Convolution

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 23/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We can essentially think of the image
as a collection of parameters
55

27
23
Keep the weights of trained convolu-
11 Dumbell
11
3

3
5
3
3
3
3
3
3
9

3
3
7

3
3
5 dense

2 dense dense
tional neural network fixed
5
11 5 2

Now adjust these parameters(image


7
256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution

pixels) so that the score of a class is


384
256
Convolution
MaxPooling 1000
256
96
Convolution

96

Convolution
MaxPooling
4096 4096
maximized
Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 23/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We can essentially think of the image
as a collection of parameters
55

27
23
Keep the weights of trained convolu-
11 Dumbell
11
3

3
5
3
3
3
3
3
3
9

3
3
7

3
3
5 dense

2 dense dense
tional neural network fixed
5
11 5 2

Now adjust these parameters(image


7
256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution

pixels) so that the score of a class is


384
256
Convolution
MaxPooling 1000
256
96
Convolution

96

Convolution
MaxPooling
4096 4096
maximized
Input

Let us see how

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 23/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
55

27
23
11
Classi
9
3 5 7
5 dense
3 3 3 3 3
Zero Image 3
5
3 3 3 3 3
2 dense dense

5 2
7
256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution
384
256
Convolution
MaxPooling 1000
256
96
Convolution
MaxPooling
4096 4096
96

Convolution

1 Start with a zero image

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 24/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
55

27
23
11
Classi
9
3 5 7
5 dense
3 3 3 3 3
Zero Image 3
5
3 3 3 3 3
2 dense dense

5 2
7
256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution
384
256
Convolution
MaxPooling 1000
256
96
Convolution
MaxPooling
4096 4096
96

Convolution

1 Start with a zero image


2 Set the score vector to be [0, 0, . . . 1, 0, 0]

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 24/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
55

27
23
11
Classi
9
3 5 7
5 dense
3 3 3 3 3
Zero Image 3
5
3 3 3 3 3
2 dense dense

5 2
7
256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution
384
256
Convolution
MaxPooling 1000
256
96
Convolution
MaxPooling
4096 4096
96

Convolution

1 Start with a zero image


2 Set the score vector to be [0, 0, . . . 1, 0, 0]
∂Sc (I)
3 Compute the gradient ∂ik

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 24/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
55

27
23
11
Classi
9
3 5 7
5 dense
3 3 3 3 3
Zero Image 3
5
3 3 3 3 3
2 dense dense

5 2
7
256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution
384
256
Convolution
MaxPooling 1000
256
96
Convolution
MaxPooling
4096 4096
96

Convolution

1 Start with a zero image


2 Set the score vector to be [0, 0, . . . 1, 0, 0]
∂Sc (I)
3 Compute the gradient ∂ik
4 Now update the pixel ik = ik − η ∂S∂ic (I)
k

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 24/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
55

27
23
11
Classi
9
3 5 7
5 dense
3 3 3 3 3
Zero Image 3
5
3 3 3 3 3
2 dense dense

5 2
7
256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution
384
256
Convolution
MaxPooling 1000
256
96
Convolution
MaxPooling
4096 4096
96

Convolution

1 Start with a zero image


2 Set the score vector to be [0, 0, . . . 1, 0, 0]
∂Sc (I)
3 Compute the gradient ∂ik
4 Now update the pixel ik = ik − η ∂S∂ic (I)
k
5 Now again do a forward pass through the network

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 24/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
55

27
23
11
Classi
9
3 5 7
5 dense
3 3 3 3 3
Zero Image 3
5
3 3 3 3 3
2 dense dense

5 2
7
256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution
384
256
Convolution
MaxPooling 1000
256
96
Convolution
MaxPooling
4096 4096
96

Convolution

1 Start with a zero image


2 Set the score vector to be [0, 0, . . . 1, 0, 0]
∂Sc (I)
3 Compute the gradient ∂ik
4 Now update the pixel ik = ik − η ∂S∂ic (I)
k
5 Now again do a forward pass through the network
6 Go to step 2

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 24/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
55

27
23
11 Dumbell
9
3 5 7
5 dense
3 3 3 3 3
3 3 2 dense dense
3 3 3
5 3
5 2
7
256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution
384
256
Convolution
MaxPooling 1000
256
96
Convolution
MaxPooling
4096 4096
96

Convolution

Lets look at the images obtained for maximizing some class scores

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 25/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
55

27
23
11
9 Cup
3 5 7
5 dense
3 3 3 3 3
3 3 2 dense dense
3 3 3
5 3
5 2
7
256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution
384
256
Convolution
MaxPooling 1000
256
96
Convolution
MaxPooling
4096 4096
96

Convolution

Lets look at the images obtained for maximizing some class scores

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 25/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
55

27
23
11
9
3 5 7
5 dense
3 3 3 3 3
3 3 2 dense dense
3 3 3
5 3
5 2
7
256
11 9 256
Dalmation
MaxPooling
27 23 384 Convolution

55 Convolution
384
256
Convolution
MaxPooling 1000
256
96
Convolution
MaxPooling
4096 4096
96

Convolution

Lets look at the images obtained for maximizing some class scores

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 25/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
55

27
23
11
9
3 5 7
5 dense
3 3 3 3 3
3 3 2 dense dense
3 3 3
5 3
5 2
7
256
11 9 256
MaxPooling
27 23 384 Convolution Bell Pepper
55 Convolution
384
256
Convolution
MaxPooling 1000
256
96
Convolution
MaxPooling
4096 4096
96

Convolution

Lets look at the images obtained for maximizing some class scores

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 25/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
55

27
23
11
9
3 5 7
5 dense
3 3 3 3 3 Lemon
3 3 2 dense dense
3 3 3
5 3
5 2
7
256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution
384
256
Convolution
MaxPooling 1000
256
96
Convolution
MaxPooling
4096 4096
96

Convolution

Lets look at the images obtained for maximizing some class scores

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 25/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
55

27
23
Husky
11
9
3 5 7
5 dense
3 3 3 3 3
3 3 2 dense dense
3 3 3
5 3
5 2
7
256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution
384
256
Convolution
MaxPooling 1000
256
96
Convolution
MaxPooling
4096 4096
96

Convolution

Lets look at the images obtained for maximizing some class scores

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 25/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
227

55

27
23
11
11 9
7

We can actually do
3 5 dense
3 3 5
3 3 3
3 3 2 dense dense
3 3 3
5 3
11 5 2
7

this for any arbit-


256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution
384
227

96
256
256

MaxPooling
Convolution
1000 rary neuron in the
Convolution
MaxPooling

96

Convolution
4096 4096

convnet
3

Input S = 2,P = 0
Parameters: 0

Repeat:
Feed an image through the network

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 26/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
227

55

27
23
11
11 9
7

We can actually do
3 5 dense
3 3 5
3 3 3
3 3 2 dense dense
3 3 3
5 3
11 5 2
7

this for any arbit-


256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution
384
227

96
256
256

MaxPooling
Convolution
1000 rary neuron in the
Convolution
MaxPooling

96

Convolution
4096 4096

convnet
3

Input S = 2,P = 0
Parameters: 0

Repeat:
Feed an image through the network
Set activation in layer of interest to all zero, except for a neuron of interest

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 26/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
227

55

27
23
11
11 9
7

We can actually do
3 5 dense
3 3 5
3 3 3
3 3 2 dense dense
3 3 3
5 3
11 5 2
7

this for any arbit-


256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution
384
227

96
256
256

MaxPooling
Convolution
1000 rary neuron in the
Convolution
MaxPooling

96

Convolution
4096 4096

convnet
3

Input S = 2,P = 0
Parameters: 0

Repeat:
Feed an image through the network
Set activation in layer of interest to all zero, except for a neuron of interest
Backprop to image

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 26/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
227

55

27
23
11
11 9
7

We can actually do
3 5 dense
3 3 5
3 3 3
3 3 2 dense dense
3 3 3
5 3
11 5 2
7

this for any arbit-


256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution
384
227

96
256
256

MaxPooling
Convolution
1000 rary neuron in the
Convolution
MaxPooling

96

Convolution
4096 4096

convnet
3

Input S = 2,P = 0
Parameters: 0

Repeat:
Feed an image through the network
Set activation in layer of interest to all zero, except for a neuron of interest
Backprop to image
ik = ik − η ∂A(I)
∂ik , A(I) is the activation of the ith neuron in some layer

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 26/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Let us look at some “updated” images
which excite certain neurons in some
layer

Layer-8

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 27/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Let us look at some “updated” images
which excite certain neurons in some
layer
Starting with different initializations
instead of using a zero image we can
get different insights

Layer-8

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 27/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Let us look at some “updated” images
which excite certain neurons in some
layer
Starting with different initializations
instead of using a zero image we can
get different insights
Each of these 4 images are obtained
Layer-8
by focusing on one neuron in layer 8
and starting with different initializa-
tions

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 27/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Let us look at some “updated” images
which excite certain neurons in some
layer
Starting with different initializations
instead of using a zero image we can
get different insights
Each of these 4 images are obtained
Layer-8
by focusing on one neuron in layer 8
and starting with different initializa-
tions
We can do a similar analysis with
other layers

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 27/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Let us look at some “updated” images
which excite certain neurons in some
layer
Starting with different initializations
instead of using a zero image we can
get different insights
Each of these 4 images are obtained
Layer-8
by focusing on one neuron in layer 8
and starting with different initializa-
tions
We can do a similar analysis with
other layers

Layer-7
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 27/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Module : Creating images from embeddings

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 28/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We could think of the fc7 layer as
some kind of an embedding for the
image
227

55

27
23
11
11 9
3 5 7
5 dense
3 3 3 3 3
3 3 2 dense dense
3 3 3
5 3
11 5 2
7
256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution
384
256
227
Convolution
MaxPooling 1000
256
96
Convolution
MaxPooling
4096 4096
96

Convolution

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 29/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We could think of the fc7 layer as
some kind of an embedding for the
image
227

55

27
23
11
11 9
3

3
5

5
3
3
3
3
3
3
3
3
7

3
3
5 dense

2 dense dense
Question: Given this embedding
11 2

can we reconstruct the image?


5
7
256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution
384
256
227
Convolution
MaxPooling 1000
256
96
Convolution
MaxPooling
4096 4096
96

Convolution

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 29/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We could think of the fc7 layer as
some kind of an embedding for the
image
227

55

27
23
11
11 9
3

3
5

5
3
3
3
3
3
3
3
3
7

3
3
5 dense

2 dense dense
Question: Given this embedding
11 2

can we reconstruct the image?


5
7
256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution

227

256
256

MaxPooling
384

Convolution
1000
We can pose this as an optimization
96

96
MaxPooling
Convolution
4096 4096
problem
Convolution

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 29/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Find an image such that
227

55

27
23
11
11 9
3 5 7
5 dense
3 3 3 3 3
3 3 2 dense dense
3 3 3
5 3
11 5 2
7
256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution
384
256
227
Convolution
MaxPooling 1000
256
96
Convolution
MaxPooling
4096 4096
96

Convolution

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 30/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Find an image such that
227
Its embedding is similar to a given
55

27
23
11
embedding
11 9
3 5 7
5 dense
3 3 3 3 3
3 3 2 dense dense
3 3 3
5 3
11 5 2
7
256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution
384
256
227
Convolution
MaxPooling 1000
256
96
Convolution
MaxPooling
4096 4096
96

Convolution

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 30/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Find an image such that
227
Its embedding is similar to a given
55

27
23
11
embedding
11 9
5 7

It looks natural (some prior regular-


3 dense
3 3 5
3 3 3
3 3 2 dense dense
3 3 3
5 3
11 5 2
7

27 23
11 9

384
256

Convolution
256

MaxPooling ization)
55 Convolution
384
256
227
Convolution
MaxPooling 1000
256
96
Convolution
MaxPooling
4096 4096
96

Convolution

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 30/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
ϕ0 :Embedding of an image of interest
227

55

27
23
11
11 9
3 5 7
5 dense
3 3 3 3 3
3 3 2 dense dense
3 3 3
5 3
11 5 2
7
256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution
384
256
227
Convolution
MaxPooling 1000
256
96
Convolution
MaxPooling
4096 4096
96

Convolution

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 31/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
ϕ0 :Embedding of an image of interest
227
X :Random image (say zero image)
55

27
23
11
11 9
3 5 7
5 dense
3 3 3 3 3
3 3 2 dense dense
3 3 3
5 3
11 5 2
7
256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution
384
256
227
Convolution
MaxPooling 1000
256
96
Convolution
MaxPooling
4096 4096
96

Convolution

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 31/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
ϕ0 :Embedding of an image of interest
227
X :Random image (say zero image)
Repeat
55

27
23
11
11 9
3 5 7
5 dense
3 3 3 3 3
3 3 2 dense dense
3 3 3
5 3
11 5 2
7
256
11 9 256
MaxPooling
27 23 384 Convolution

55 Convolution
384
256
227
Convolution
MaxPooling 1000
256
96
Convolution
MaxPooling
4096 4096
96

Convolution

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 31/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
ϕ0 :Embedding of an image of interest
227
X :Random image (say zero image)
Repeat
55

27
23
11
11 9
7

Forward pass using X and compute


3 5 dense
3 3 5
3 3 3
3 3 2 dense dense
3 3 3
5 3
11 5 2

23
11 9

384
7
256
256

MaxPooling
ϕ(x).
27 Convolution

55 Convolution
384
256
227
Convolution
MaxPooling 1000
256
96
Convolution
MaxPooling
4096 4096
96

Convolution

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 31/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
ϕ0 :Embedding of an image of interest
227
X :Random image (say zero image)
Repeat
55

27
23
11
11 9
7

Forward pass using X and compute


3 5 dense
3 3 5
3 3 3
3 3 2 dense dense
3 3 3
5 3
11 5 2

23
11 9

384
7
256
256

MaxPooling
ϕ(x).
27 Convolution

227
55
256
384

Convolution
Convolution
Compute
MaxPooling 1000
256
96
Convolution

L (i) = ||ϕ(x) − ϕ0 ||2 + λ||ϕ(x)||66


MaxPooling
4096 4096
96

Convolution

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 31/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
ϕ0 :Embedding of an image of interest
227
X :Random image (say zero image)
Repeat
55

27
23
11
11 9
7

Forward pass using X and compute


3 5 dense
3 3 5
3 3 3
3 3 2 dense dense
3 3 3
5 3
11 5 2

23
11 9

384
7
256
256

MaxPooling
ϕ(x).
27 Convolution

227
55
256
384

Convolution
Convolution
Compute
MaxPooling 1000
256
96
Convolution

L (i) = ||ϕ(x) − ϕ0 ||2 + λ||ϕ(x)||66


MaxPooling
4096 4096
96

Convolution

Input

ik = ik − η L∂i(i)
k

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 31/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Original Image Conv-1

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 32/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Original Image Relu-1

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 32/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Original Image Mpool-1

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 32/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Original Image Norm-1

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 32/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Original Image Conv-2

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 32/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Original Image Relu-2

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 32/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Original Image Mpool-2

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 32/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Original Image Norm-2

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 32/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Original Image Conv-3

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 32/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Original Image Relu-3

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 32/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Original Image Conv-4

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 32/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Original Image Relu-4

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 32/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Original Image Conv-5

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 32/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Original Image Relu-5

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 32/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Original Image Mpool-5

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 32/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Original Image FC-6

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 32/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Original Image Relu-6

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 32/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Original Image FC-7

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 32/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Original Image Relu-7

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 32/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Original Image FC-8

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 32/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Module : Deep Dream

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 33/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Suppose instead of starting with a
227
blank (zero) image we start with an
55

27
23
11
actual image.
11 9
3 7
3 5
3 3 3
3 2
3 3 3 3
5 2
11
7 256
9 256
11
MaxPooling
23 384
27 Convolution

55 384 Convolution
256
227
Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

Convolution
3

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 34/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Suppose instead of starting with a
227
blank (zero) image we start with an
55

27
23
11
actual image.
11 9
3 7
3 5
3 3 3
3 2
3 3 3 3
5 2
11
7 256
9 256
11
MaxPooling
23 384
27 Convolution

55 384 Convolution
256
227
Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

Convolution
3

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 34/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Suppose instead of starting with a
227
blank (zero) image we start with an
55

27
23
11
actual image.
11 9
3 7

3
3

3
3

3
3

3
3

3
5

2
2 We focus on some layer and check the
11 5

27 23
11 9
384
7
256
256

MaxPooling
activations of the neurons
Convolution

55 384 Convolution
256
227
Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

Convolution
3

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 34/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Suppose instead of starting with a
227
blank (zero) image we start with an
55

27
23
11
actual image.
11 9
3 7

3
3

3
3

3
3

3
3

3
5

2
2 We focus on some layer and check the
11 5

27 23
11 9
384
7
256
256

MaxPooling
activations of the neurons
Convolution

384 Convolution

We want to change the image so that


55
256
227
Convolution
MaxPooling
256
96

96
MaxPooling
Convolution
these neurons fire even more
Convolution
3

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 34/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
How would we achieve this?
227

55

27
23
11
11 9
3 7
3 5
3 3 3
3 2
3 3 3 3
5 2
11
7 256
9 256
11
MaxPooling
23 384
27 Convolution

55 384 Convolution
256
227
Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

Convolution
3

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 35/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
How would we achieve this?
227

55
Suppose we want to boost the activa-
11
27
23
11
9
tion hij (some neuron in some layer)
3 7
3 5
3 3 3
3 2
3 3 3 3
5 2
11
7 256
9 256
11
MaxPooling
23 384
27 Convolution

55 384 Convolution
256
227
Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

Convolution
3

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 35/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
How would we achieve this?
227
hij Suppose we want to boost the activa-
55

11
27
23
11
9
tion hij (some neuron in some layer)
3 7
3 5
3 3 3
3 2
3 3 3 3
5 2
11
7 256

27 23
11 9
384
256

Convolution
MaxPooling We can formulate this as the following
384 Convolution

optimization problem
55
256
227
Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

Convolution
3

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 35/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
How would we achieve this?
227
hij Suppose we want to boost the activa-
55

11
27
23
11
9
tion hij (some neuron in some layer)
3 7
3 5
3 3 3
3 2
3 3 3 3
5 2
11
7 256

27 23
11 9
384
256

Convolution
MaxPooling We can formulate this as the following
384 Convolution

optimization problem
55
256
227
Convolution
MaxPooling
256
96

Convolution
MaxPooling

max L (I)
96

Convolution
3

Input
I
L (I) = h2ij

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 35/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
How would we achieve this?
227
hij Suppose we want to boost the activa-
55

11
27
23
11
9
tion hij (some neuron in some layer)
3 7
3 5
3 3 3
3 2
3 3 3 3
5 2
11
7 256

27 23
11 9
384
256

Convolution
MaxPooling We can formulate this as the following
384 Convolution

optimization problem
55
256
227
Convolution
MaxPooling
256
96

Convolution
MaxPooling

max L (I)
96

Convolution
3

Input
I
L (I) = h2ij

Consider a pixel imn in the image

∂L (I) ∂L (I) ∂hij


=
∂imn ∂hij ∂imn

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 35/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
(
Once the image is updated imn =
227
)
imn + ∂L (I)
we feed it back to the
55

27

11
23
11
9
∂imn
3
3
3

3
3

3
3

3
7

3
5
2
network
3
5 2
11
7 256
9 256
11
MaxPooling
23 384
27 Convolution

55 384 Convolution
256
227
Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

Convolution
3

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 36/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
(
hij Once the image is updated imn =
227
)
imn + ∂L (I)
we feed it back to the
55

27

11
23
11
9
∂imn
3
3
3

3
3

3
3

3
7

3
5
2
network
3
5 2
11
256

This time the target neurons should


7
9 256
11
MaxPooling
23 384
27 Convolution

227
55
256
384

Convolution
Convolution
fire even more (because we have pre-
MaxPooling
256

96
96

MaxPooling
Convolution cisely modified the image to achieve
3
Convolution this)
Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 36/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
(
hij Once the image is updated imn =
227
)
imn + ∂L (I)
we feed it back to the
55

27

11
23
11
9
∂imn
3
3
3

3
3

3
3

3
7

3
5
2
network
3
5 2
11
256

This time the target neurons should


7
9 256
11
MaxPooling
23 384
27 Convolution

227
55
256
384

Convolution
Convolution
fire even more (because we have pre-
MaxPooling
256

96
96

MaxPooling
Convolution cisely modified the image to achieve
3
Convolution this)
Input

Doing this iteratively would make the


image more and more like the pat-
terns that cause the neuron to fire

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 36/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
(
hij Once the image is updated imn =
227
)
imn + ∂L (I)
we feed it back to the
55

27

11
23
11
9
∂imn
3
3
3

3
3

3
3

3
7

3
5
2
network
3
5 2
11
256

This time the target neurons should


7
9 256
11
MaxPooling
23 384
27 Convolution

227
55
256
384

Convolution
Convolution
fire even more (because we have pre-
MaxPooling
256

96
96

MaxPooling
Convolution cisely modified the image to achieve
3
Convolution this)
Input

Doing this iteratively would make the


image more and more like the pat-
terns that cause the neuron to fire
Let us run this algorithm

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 36/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 37/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 38/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 39/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
So what exactly is happening here?
227

55

27
23
11
11 9
3 7
3 5
3 3 3
3 2
3 3 3 3
5 2
11
7 256
9 256
11
MaxPooling
23 384
27 Convolution

55 384 Convolution
256
227
Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

Convolution
3

Input


research.googleblog.com/2015/06/inceptionism-
. . . . . . . . . . . . . . . . . . . .
going-deeper-into-neural.html
. . . . . . . . . . . . . . . . . . . . 40/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
So what exactly is happening here?
227

55
The network has been trained to detect
11
27
23
11
9
certain patterns (dogs, cat, birds etc.)
3 7

3
3

3
3

3
3

3
3

3
5

2
2 which appear frequently in the ImageNet
11 5

27 23
11 9
384
7
256
256

MaxPooling
data
Convolution

55 384 Convolution
256
227
Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

Convolution
3

Input


research.googleblog.com/2015/06/inceptionism-
. . . . . . . . . . . . . . . . . . . .
going-deeper-into-neural.html
. . . . . . . . . . . . . . . . . . . . 40/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
So what exactly is happening here?
227

55
The network has been trained to detect
11
27
23
11
9
certain patterns (dogs, cat, birds etc.)
3 7

3
3

3
3

3
3

3
3

3
5

2
2 which appear frequently in the ImageNet
11 5

27 23
11 9
384
7
256
256

MaxPooling
data
Convolution

384 Convolution

It starts seeing these patterns even when


55
256
227
Convolution
MaxPooling
256
96

96
MaxPooling
Convolution
they hardly exist
Convolution
3

Input


research.googleblog.com/2015/06/inceptionism-
. . . . . . . . . . . . . . . . . . . .
going-deeper-into-neural.html
. . . . . . . . . . . . . . . . . . . . 40/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
So what exactly is happening here?
227

55
The network has been trained to detect
11
27
23
11
9
certain patterns (dogs, cat, birds etc.)
3 7

3
3

3
3

3
3

3
3

3
5

2
2 which appear frequently in the ImageNet
11 5

27 23
11 9
384
7
256
256

MaxPooling
data
Convolution

384 Convolution

It starts seeing these patterns even when


55
256
227
Convolution
MaxPooling
256
96

96
MaxPooling
Convolution
they hardly exist
If a cloud looks a little bit like a bird,
Convolution
3

Input

the network will make it look more like


a bird. This in turn will make the net-
work recognize the bird even more strongly
on the next pass and so forth, until a
highly detailed bird appears seemingly out
of nowhere. - Google∗


research.googleblog.com/2015/06/inceptionism-
. . . . . . . . . . . . . . . . . . . .
going-deeper-into-neural.html
. . . . . . . . . . . . . . . . . . . . 40/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Module : Deep Art

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 41/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 42/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 42/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 42/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 42/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 42/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
To design a network which can do
this, we first define two quantities

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 43/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
To design a network which can do
227

55

27
23
this, we first define two quantities
11
11 9

3
3
3

3
3

3
3

3
7

3
5
5

2
2 Content Targets : The activations
11
7 256

55
27 23
11

384
9
384

Convolution
256

Convolution
MaxPooling
of all layers for the given content im-
256
227

96
256

Convolution
MaxPooling
Convolution
age
MaxPooling
96

Convolution
3

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 43/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
To design a network which can do
227

55

27
23
this, we first define two quantities
11
11 9

3
3

3
3

3
3 3

3
7

3
5
5

2
2 Content Targets : The activations
11
7 256

55
27 23
11

384
9
384

Convolution
256

Convolution
MaxPooling
of all layers for the given content im-
256
227

96
256

Convolution
MaxPooling
Convolution
age
MaxPooling
96

3
Convolution
Ideally, we would want the new im-
Input

227

55
age to be such that
27
23
11
11 9
3 7
3 5
3 3 3
3 2
3 3 3 3
5 2
11
7 256
9 256
11
MaxPooling
23 384
27 Convolution

55 384 Convolution
256
227
Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

Convolution
3

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 43/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
To design a network which can do
227

55

27
23
this, we first define two quantities
11
11 9

3
3

3
3

3
3 3

3
7

3
5
5

2
2 Content Targets : The activations
11
7 256

55
27 23
11

384
9
384

Convolution
256

Convolution
MaxPooling
of all layers for the given content im-
256
227

96
256

Convolution
MaxPooling
Convolution
age
MaxPooling
96

3
Convolution
want them to be equal Ideally, we would want the new im-
Input

227

55
age to be such that it’s activations are
11
3
27
23
11
9
7
also close to those of the original con-
3 5
3

tent image
3 3
3 2
3 3 3 3
5 2
11
7 256
9 256
11
MaxPooling
23 384
27 Convolution

55 384 Convolution
256
227
Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

Convolution
3

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 43/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
To design a network which can do
227

55

27
23
this, we first define two quantities
11
11 9

3
3

3
3

3
3 3

3
7

3
5
5

2
2 Content Targets : The activations
11
7 256

55
27 23
11

384
9
384

Convolution
256

Convolution
MaxPooling
of all layers for the given content im-
256
227

96
256

Convolution
MaxPooling
Convolution
age
MaxPooling
96

3
Convolution
want them to be equal Ideally, we would want the new im-
Input

227

55
age to be such that it’s activations are
11
3
27
23
11
9
7
also close to those of the original con-
3 5
3

tent image
3 3
3 2
3 3 3 3
5 2
11
7 256
9 256
11
MaxPooling
23 384
27 Convolution

227
55
256

MaxPooling
384

Convolution
Convolution
Let p⃗, ⃗x be the activations of the con-
256
96

96
MaxPooling
Convolution
tent image and the new image (to be
Convolution
3

Input generated) respectively



Lcontent (⃗
p, ⃗x) = pijk − ⃗xijk )2
(⃗
ijk
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 43/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Next we would want the style of the
227

55

27
23
11
generated image to be the same as the
11 9

3
3
3

3
3

3
3

3
7

3
5

2
2
style image
11 5
7 256
9 256
11
MaxPooling
23 384
27 Convolution

55 384 Convolution
256
227
Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

Convolution
3

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 44/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Next we would want the style of the
227

55

27
23
11
generated image to be the same as the
11 9

3
3
3

3
3

3
3

3
7

3
5

2
2
style image
11 5
7 256
256

How do we capture the style of the


11 9
MaxPooling
23 384
27 Convolution

55 384 Convolution
256
227

image?
Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

Convolution
3

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 44/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Next we would want the style of the
227

55

27
23
11
generated image to be the same as the
11 9

3
3
3

3
3

3
3

3
7

3
5

2
2
style image
11 5
7 256
256

How do we capture the style of the


11 9
MaxPooling
23 384
27 Convolution

55 384 Convolution
256
227

image?
Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

Input
Convolution

Turns out that if V ∈ R64×(256×256) is


the activation at a layer then V T V ∈
R64×64 captures the style of the image

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 44/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Next we would want the style of the
227

55

27
23
11
generated image to be the same as the
11 9

3
3
3

3
3

3
3

3
7

3
5

2
2
style image
11 5
7 256
256

How do we capture the style of the


11 9
MaxPooling
23 384
27 Convolution

55 384 Convolution
256
227

image?
Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

Input
Convolution

Turns out that if V ∈ R64×(256×256) is


the activation at a layer then V T V ∈
R64×64 captures the style of the image

The deeper layers capture more of


this style information

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 44/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Next we would want the style of the
227

55

27
23
11
generated image to be the same as the
11 9

3
3
3

3
3

3
3

3
7

3
5

2
2
style image
11 5
7 256
256

How do we capture the style of the


11 9
MaxPooling
23 384
27 Convolution

55 384 Convolution
256
227

image?
Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

Input
Convolution

Turns out that if V ∈ R64×(256×256) is


the activation at a layer then V T V ∈
R64×64 captures the style of the image

The deeper layers capture more of


this style information

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 44/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Next we would want the style of the
227

55

27
23
11
generated image to be the same as the
11 9

3
3
3

3
3

3
3

3
7

3
5

2
2
style image
11 5
7 256
256

How do we capture the style of the


11 9
MaxPooling
23 384
27 Convolution

55 384 Convolution
256
227

image?
Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

Input
Convolution

Turns out that if V ∈ R64×(256×256) is


the activation at a layer then V T V ∈
R64×64 captures the style of the image

The deeper layers capture more of


this style information

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 44/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Next we would want the style of the
227

55

27
23
11
generated image to be the same as the
11 9

3
3
3

3
3

3
3

3
7

3
5

2
2
style image
11 5
7 256
256

How do we capture the style of the


11 9
MaxPooling
23 384
27 Convolution

55 384 Convolution
256
227

image?
Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

Input
Convolution

Turns out that if V ∈ R64×(256×256) is


the activation at a layer then V T V ∈
R64×64 captures the style of the image

The deeper layers capture more of


this style information

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 44/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
To ensure that the style of the new
image captured by layer ℓ matches
the style of the style image, we can
227

55

27
23
11

use the following objective function :


11 9
3 7
3 5
3 3 3
3 2
3 3 3 3
5 2
11
7 256
9 256
11
MaxPooling
23 384
27 Convolution

55 384 Convolution
256
227
Convolution
MaxPooling
256

Content Image
96

Convolution
MaxPooling
96

Convolution
3

Input

227

55

27
23
11
11 9
3 7
3 5
3 3 3
3 2
3 3 3 3
5 2
11
7 256
9 256
11
MaxPooling
23 384
27 Convolution

55 384 Convolution
256
227
Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

Convolution
3

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 45/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
To ensure that the style of the new
image captured by layer ℓ matches
the style of the style image, we can
227

55

27
23
11

use the following objective function :


11 9
3 7
3 5
3 3 3
3 2
3 3 3 3
5 2
11
7 256
9 256
11
MaxPooling
23 384
27 Convolution


55 384 Convolution
256
227

V2T V2
Convolution
MaxPooling
256

Content Image (Gℓj − Aℓij )2


96

96
MaxPooling
Convolution

Eℓ =
3
Convolution
equal
Input ij
V1T V1
227

55

where Gℓ and Aℓ are the style gram


27
23
11
11 9
3 7
3 5
3 3 3
3 2

11
3

11
3

9
384
3

7
256
3
5
256

MaxPooling
2
matrices computed at layer ℓ for the
27 23 Convolution

227
55
256

MaxPooling
384

Convolution
Convolution
style image and new image respect-
256
96

96
MaxPooling
Convolution
ively.
Convolution
3

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 45/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
To ensure that the style of the new
image captured by layer ℓ matches
the style of the style image, we can
227

55

27
23
11

use the following objective function :


11 9
3 7
3 5
3 3 3
3 2
3 3 3 3
5 2
11
7 256
9 256
11
MaxPooling
23 384
27 Convolution


55 384 Convolution
256
227

V2T V2
Convolution
MaxPooling
256

Content Image (Gℓj − Aℓij )2


96

96
MaxPooling
Convolution

Eℓ =
3
Convolution
equal
Input ij
V1T V1
227

55

where Gℓ and Aℓ are the style gram


27
23
11
11 9
3 7
3 5
3 3 3
3 2

11
3

11
3

9
384
3

7
256
3
5
256

MaxPooling
2
matrices computed at layer ℓ for the
27 23 Convolution

227
55
256

MaxPooling
384

Convolution
Convolution
style image and new image respect-
256
96

96
MaxPooling
Convolution
ively.

L
Convolution
3

Input

Lstyle (⃗a, x̄) = wℓ Eℓ


ℓ=0
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 45/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
227 227

55 55

27 27
23 23
11 11
11 9 11 9
3 5 7 3 5 7
3 3 5 3 3 5
3 3 3 3 3 3
3 2 3 2
3 3 3 3 3 3
5 3 3 5 3 3
5 2 5 2
11 11
7 256 7 256
9 256 9 256
11 11
MaxPooling MaxPooling
23 384 23 384
27 Convolution 27 Convolution

55 384 Convolution 55 384 Convolution


256 256
227 227
Convolution Convolution
MaxPooling MaxPooling
256 256
96 96

Convolution Convolution
MaxPooling MaxPooling
96 96

Convolution Convolution
3 3

Input Input

227

55

27
23
11
11 9
3 5 7
3 3 5
3 3 3
3 2
3 3 3
5 3 3
5 2
11
7 256
9 256
11
MaxPooling
23 384
27 Convolution

55 384 Convolution
256
227
Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

Convolution
3

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 46/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
227 227

55 55

27 27
23 23
11 11
11 9 11 9
3 7 3 7
3 5 3 5
3 3 3 3 3 3
3 2 3 2
3 3 3 3 3 3
3 3
5 2 5 2
11 11
7 256 7 256
9 256 9 256
11 11
MaxPooling MaxPooling
23 384 23 384
27 Convolution 27 Convolution

55 384 Convolution 55 384 Convolution


256 256
227 227
Convolution Convolution
MaxPooling MaxPooling
256 256
96 96

Convolution Convolution
MaxPooling MaxPooling
96 96

Convolution Convolution
3 3

Input Input

227

55

27
23
11
11 9
3 7
3 5
3 3 3
3 2
3 3 3 3
5 2
11
7 256
9 256
11
MaxPooling
23 384
27 Convolution

55 384 Convolution
256
227
Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

Convolution
3

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 46/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
227 227

55 55

27 27
23 23
11 11
11 9 11 9
3 7 3 7
3 5 3 5
3 3 3 3 3 3
3 2 3 2
3 3 3 3 3 3
3 3
5 2 5 2
11 11
7 256 7 256
9 256 9 256
11 11
MaxPooling MaxPooling
23 384 23 384
27 Convolution 27 Convolution

55 384 Convolution 55 384 Convolution


256 256
227 227
Convolution Convolution
MaxPooling MaxPooling
256 256
96 96

Convolution Convolution
MaxPooling MaxPooling
96 96

Convolution Convolution
3 3

Input Input

227

55

27
23
11
11 9
3 7
3 5
3 3 3
3 2
3 3 3 3
5 2
11
7 256
9 256
11
MaxPooling
23 384
27 Convolution

55 384 Convolution
256
227
Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

Convolution
3

Input

The total loss is given by :-


Ltotal (⃗
p, ⃗a, ⃗x) = αLcontent (⃗
p, ⃗x) + βL.style (⃗a, ⃗x)
. . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 46/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Module : Fooling Deep Convolution Neural Networks

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 47/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Turns out that using this idea of optimizing over the input, we can also “fool”
ConvNets

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 48/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Turns out that using this idea of optimizing over the input, we can also “fool”
ConvNets
Let us see how

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 48/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Suppose we feed in an image to a
227

55

27
23
11
Convnet.
11 9
3 5 7
3 3 5
3 3 3
3 2
3 3 3
5 3 3
5 2
11
7 256
9 256
11
MaxPooling
23 384
27 Convolution

55 384 Convolution
256
227
Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

Convolution
3

Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 49/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Suppose we feed in an image to a
227
ostrich
55

27
23
11 .. Convnet.
.
11 9
3 5 7

11
3
5
3

3
3

3
3

3
3

7
3

3
5
5

256
2
2
Now instead of maximizing the log-
256 bus
.. likelihood of the correct class (bus) we
11 9
MaxPooling
23 384
27 Convolution

.
55 384 Convolution
256
227

set the objective to maximize some


Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

3
Convolution
incorrect class (say, ostrich)
Input

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 49/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Suppose we feed in an image to a
227
ostrich
55

27
23
11 .. Convnet.
.
11 9
3 5 7

11
3
5
3

3
3

3
3

3
3

7
3

3
5
5

256
2
2
Now instead of maximizing the log-
256 bus
.. likelihood of the correct class (bus) we
11 9
MaxPooling
23 384
27 Convolution

.
55 384 Convolution
256
227

set the objective to maximize some


Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

3
Convolution
incorrect class (say, ostrich)
Input

Turns out that with minimal changes


to the image (using backprop) we can
soon convince the Convnet that this
is an ostrich.

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 49/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Suppose we feed in an image to a
227
ostrich
55

27
23
11 .. Convnet.
.
11 9
3 5 7

11
3
5
3

3
3

3
3

3
3

7
3

3
5
5

256
2
2
Now instead of maximizing the log-
256 bus
.. likelihood of the correct class (bus) we
11 9
MaxPooling
23 384
27 Convolution

.
55 384 Convolution
256
227

set the objective to maximize some


Convolution
MaxPooling
256
96

Convolution
MaxPooling
96

3
Convolution
incorrect class (say, ostrich)
Input

Turns out that with minimal changes


to the image (using backprop) we can
soon convince the Convnet that this
is an ostrich.
Let us see some examples

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 49/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Notice that the changes
are so minimal that the
two images are indistin-
guishable to humans


Intriguing properties of neural networks, Szegedy et al., 2013 . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 50/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Notice that the changes
are so minimal that the
two images are indistin-
guishable to humans
But the ConvNet thinks
that the third image ob-
tained by adding the first
image to the second im-
age is an ostrich


Intriguing properties of neural networks, Szegedy et al., 2013 . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 50/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Notice that the changes
are so minimal that the
two images are indistin-
guishable to humans
But the ConvNet thinks
that the third image ob-
tained by adding the first
image to the second im-
age is an ostrich


Intriguing properties of neural networks, Szegedy et al., 2013 . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 50/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Notice that the changes
are so minimal that the
two images are indistin-
guishable to humans
But the ConvNet thinks
that the third image ob-
tained by adding the first
image to the second im-
age is an ostrich


Intriguing properties of neural networks, Szegedy et al., 2013 . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 50/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Notice that the changes
are so minimal that the
two images are indistin-
guishable to humans
But the ConvNet thinks
that the third image ob-
tained by adding the first
image to the second im-
age is an ostrich


Intriguing properties of neural networks, Szegedy et al., 2013 . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 50/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Notice that the changes
are so minimal that the
two images are indistin-
guishable to humans
But the ConvNet thinks
that the third image ob-
tained by adding the first
image to the second im-
age is an ostrich


Intriguing properties of neural networks, Szegedy et al., 2013 . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 50/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We can also do this start-
ing with random images and
then optimizing them to pre-
dict some class.


Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable
Images Nguyen, Yosinski, Clune, 2014

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 51/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We can also do this start-
ing with random images and
then optimizing them to pre-
dict some class.
In all these cases the classi-
fier is 99.6% confident of the
class


Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable
Images Nguyen, Yosinski, Clune, 2014

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 51/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We can also do this start-
ing with random images and
then optimizing them to pre-
dict some class.
In all these cases the classi-
fier is 99.6% confident of the
class


Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable
Images Nguyen, Yosinski, Clune, 2014

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 51/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We can also do this start-
ing with random images and
then optimizing them to pre-
dict some class.
In all these cases the classi-
fier is 99.6% confident of the
class


Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable
Images Nguyen, Yosinski, Clune, 2014

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 51/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We can also do this start-
ing with random images and
then optimizing them to pre-
dict some class.
In all these cases the classi-
fier is 99.6% confident of the
class


Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable
Images Nguyen, Yosinski, Clune, 2014

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 51/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We can also do this start-
ing with random images and
then optimizing them to pre-
dict some class.
In all these cases the classi-
fier is 99.6% confident of the
class


Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable
Images Nguyen, Yosinski, Clune, 2014

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 51/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We can also do this start-
ing with random images and
then optimizing them to pre-
dict some class.
In all these cases the classi-
fier is 99.6% confident of the
class


Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable
Images Nguyen, Yosinski, Clune, 2014

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 51/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We can also do this start-
ing with random images and
then optimizing them to pre-
dict some class.
In all these cases the classi-
fier is 99.6% confident of the
class


Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable
Images Nguyen, Yosinski, Clune, 2014

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 51/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
We can also do this start-
ing with random images and
then optimizing them to pre-
dict some class.
In all these cases the classi-
fier is 99.6% confident of the
class
Let us see an intuitive ex-
planation of why this hap-
pens


Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable
Images Nguyen, Yosinski, Clune, 2014

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 51/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Images are extremely high dimen-
sional objects (R227×227 )

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 52/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Images are extremely high dimen-
sional objects (R227×227 )
There are many many many points in
this high dimensional space

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 52/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Images are extremely high dimen-
sional objects (R227×227 )
There are many many many points in
this high dimensional space
Of these only a few are images (of
which we see some during training)

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 52/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Images are extremely high dimen-
sional objects (R227×227 )
There are many many many points in
this high dimensional space
Of these only a few are images (of
which we see some during training)
Using these training images we fit
some decision boundaries

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 52/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Images are extremely high dimen-
sional objects (R227×227 )
There are many many many points in
this high dimensional space
Of these only a few are images (of
which we see some during training)
Using these training images we fit
some decision boundaries
While doing so we also end up taking
decisions about the many many un-
seen points in this high dimensional
space (Notice the large green and
red regions which do not contain any
training points)

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 52/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Images are extremely high dimen-
sional objects (R227×227 )
There are many many many points in
this high dimensional space
Of these only a few are images (of
which we see some during training)
Using these training images we fit
some decision boundaries
While doing so we also end up taking
decisions about the many many un-
seen points in this high dimensional
space (Notice the large green and
red regions which do not contain any
training points)

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 52/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12
Images are extremely high dimen-
sional objects (R227×227 )
There are many many many points in
this high dimensional space
Of these only a few are images (of
which we see some during training)
Using these training images we fit
some decision boundaries
While doing so we also end up taking
decisions about the many many un-
seen points in this high dimensional
space (Notice the large green and
red regions which do not contain any
training points)

. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . 52/1
Mitesh M. Khapra CS7015 (Deep Learning) : Lecture 12

You might also like