ݺߣ

ݺߣShare a Scribd company logo
KRZYSZTOF KUDRYŃSKI
BŁAŻEJ KUBIAK
DEEP LEARNING IN COMPUTER VISION
Plan
• Get excited (7 min)
about the world we live in
• Understand CNN (18 min)
from fundamentals to state-of-the art approach
• Apply CNN (15 min)
by walk-through a challenging computer-vision project.
• Get scared (1 min)
about the world we’ll live in
Deep Learning in Computer Vision
Krzysztof Kudryński
(1982)
www.youtube.com/watch?v=qHepKd38pr0Deep Learning in Computer Vision
Krzysztof Kudryński
Deep Learning in Computer Vision
Krzysztof Kudryński
(2017)
Deep Learning in Computer Vision
Krzysztof Kudryński
(2017)
www.youtube.com/watch?v=W6wzUMk9w-cDeep Learning in Computer Vision
Krzysztof Kudryński
(2008)
Deep Learning in Computer Vision
Krzysztof Kudryński
(2016)
www.youtube.com/watch?v=FKXOucXB4a8Deep Learning in Computer Vision
ⓒKrzysztof Kudryński
www.youtube.com
(1982)
Deep Learning in Computer Vision
Krzysztof Kudryński
(2015)
Deep Learning in Computer Vision
Krzysztof Kudryński
(2015)
Deep Learning in Computer Vision
Krzysztof Kudryński
(2015)
Deep Learning in Computer Vision
Krzysztof Kudryński
A small bird with yellow
breast, brown crown and
black superciliary
(2015)
Deep Learning in Computer Vision
Krzysztof Kudryński
Deep Learning in Computer Vision
Krzysztof Kudryński
(2015-2018)
Deep Learning in Computer Vision
Krzysztof Kudryński
Deep Learning in Computer Vision
Krzysztof Kudryński
Deep Learning in Computer Vision
Krzysztof Kudryński
100%
0% 0%
Deep Learning in Computer Vision
Krzysztof Kudryński
vertical
edges
horizon
edges
frame
glasses
23
345
3%
97%
Convolution
Max-pooling
Deep Learning in Computer Vision
Krzysztof Kudryński
[n, 200, 180, 4][5, 5, 3, 4]
n
[n, 200, 180, 3]
200
180
n
3
4
3
5
5
4
200
180
DeepLearninginComputerVision
KrzysztofKudryński
vertical
edges
horizon
edges
frame
glasses
23
345
3%
97%
[n, 200, 180, 1] [3, 3, 1, 4] [n, 200, 180, 4]
[n, 20, 18, 4] [5, 5, 4, 1] [n, 20, 18, 1]
[n, 10, 9, 1] [6, 6, 1, 1] [n, 10, 9, 1]
[n, 5, 5, 1]
[25, 2] [n, 2] [n, 2]
placeholder
session
Parse
images with
labels
[0 1]
[0 1]
[1 0]
[1 0]
Deep Learning in Computer Vision
Krzysztof Kudryński
0
Deep Learning in Computer Vision
Krzysztof Kudryński
97% glasses99% glasses
51% glasses
DeepLearninginComputerVision
KrzysztofKudryński
0a
Deep Learning in Computer Vision
Krzysztof Kudryński
0b
Deep Learning in Computer Vision
Krzysztof Kudryński
1
Deep Learning in Computer Vision
Krzysztof Kudryński
1
Deep Learning in Computer Vision
Krzysztof Kudryński
Citations: 3000
ALWAYS
POSITIVE
Deep Learning in Computer Vision
Krzysztof Kudryński
1b
Deep Learning in Computer Vision
Krzysztof Kudryński
1c
Deep Learning in Computer Vision
Krzysztof Kudryński
2
Deep Learning in Computer Vision
Krzysztof Kudryński
2
Deep Learning in Computer Vision
Krzysztof Kudryński
3
Deep Learning in Computer Vision
Krzysztof Kudryński
4
LADY #0
LADY #1
NO
GLASSES
HAS
GLASSES
Deep Learning in Computer Vision
Krzysztof Kudryński
6a (5)
correct
correct
correct
correct
correct
wrong
wrong
Deep Learning in Computer Vision
Krzysztof Kudryński
6b (6)
Deep Learning in Computer Vision
Krzysztof Kudryński
7a
Deep Learning in Computer Vision
Krzysztof Kudryński
Deep Learning in Computer Vision
Krzysztof Kudryński
)
Deep Learning in Computer Vision
Krzysztof Kudryński
7b
Deep Learning in Computer Vision
Krzysztof Kudryński
We need to
cut the costs!
Optimizer().
minimize(cost)
Deep Learning in Computer Vision
Krzysztof Kudryński
We need to
cut the costs!
Optimizer().
minimize(cost)
Deep Learning in Computer Vision
Krzysztof Kudryński
)
We need to
cut the costs!
Optimizer().
minimize(cost)
Deep Learning in Computer Vision
Krzysztof Kudryński
Deep Learning in Computer Vision
Krzysztof Kudryński
7e
Deep Learning in Computer Vision
Krzysztof Kudryński
8b
Deep Learning in Computer Vision
Krzysztof Kudryński
• Smart initialization
• Regularization
• Gradient clipping
• New activation functions
• Huge training sets
• Deep networks
Sophisticated
problems
Clever solutions
• Hard to train
• Vanishing gradient
• Exploding gradient
Deep Learning in Computer Vision
Krzysztof Kudryński
• Smart initialization
• Regularization
• Gradient clipping
• New activation functions
• Intermediate training paths
• Huge training sets
• Deep networks
• Hard to train
• Vanishing gradient
• Exploding gradient
Sophisticated
problems
Clever solutions
Deep Learning in Computer Vision
Krzysztof Kudryński
• Smart initialization
• Regularization
• Gradient clipping
• New activation functions
• Intermediate training paths
• Residual connections
• Huge training sets
• Deep networks
• Hard to train
• Vanishing gradient
• Exploding gradient
Sophisticated
problems
Clever solutions
Deep Learning in Computer Vision
Krzysztof Kudryński
Deep Learning in Computer Vision
Krzysztof Kudryński
POLAND /112018
Deep Learning in Computer Vision
Krzysztof Kudryński
Deep Learning in Computer Vision
Krzysztof Kudryński
Deep Learning in Computer Vision
Krzysztof Kudryński
Deep Learning in Computer Vision
Krzysztof Kudryński
RGB Image
Convolution
Max pooling
X X
Convolution
Max pooling
A
B
C
||A – B||
||B – C||
L = ||A – Ex||2
L = ||B – C||2 – ||A – B||2
Deep Learning in Computer Vision
Krzysztof Kudryński
Deep Learning in Computer Vision
Krzysztof Kudryński
Convolution
Max pooling
X X
Convolution
Max pooling
A
B
|A – B|
Face recognition network
Deep Learning in Computer Vision
Krzysztof Kudryński
Deep Learning in Computer Vision
Krzysztof Kudryński
Kate Winslet
Arnold Schwarzenegger
Alec Baldwin
Abdullah
Deep Learning in Computer Vision
Krzysztof Kudryński
Convolution
Max pooling
X X
Convolution
Max pooling
A
B
|A – B|
Face recognition network
1
Deep Learning in Computer Vision
Krzysztof Kudryński
Convolution
Max pooling
X X
Convolution
Max pooling
A
B
|A – B|
Face recognition network
0
Deep Learning in Computer Vision
Krzysztof Kudryński
model = Sequential()
model.add(Dense(rowSize, input_dim=rowSize, activation='relu'))
model.add(Dense(rowSize, activation='relu'))
model.add(Dense(1, activation='sigmoid'))
adam = optimizers.adam()
model.compile(loss='binary_crossentropy',
optimizer=adam,
metrics=['accuracy'])
model.fit_generator(generate_arrays_from_file(trainArr, tranLabelMap, 30),
steps_per_epoch=2000, epochs=100,
validation_data=(valData, valDataLabels))
Training Network in Keras
Deep Learning in Computer Vision
Krzysztof Kudryński
(2018)
source:http://dx.doi.org/10.1101/272518
Deep Learning in Computer Vision
Krzysztof Kudryński
(2018)
source:http://dx.doi.org/10.1101/272518
Deep Learning in Computer Vision
Krzysztof Kudryński
Krzysztof Kudrynski
k.kudrynski@gmail.com
@kriskudrynski
Blazej Kubiak
blazejkubiak@gmail.com
@blazkubiak
Deep Learning in Computer Vision
Krzysztof Kudryński

More Related Content

infoShare AI Roadshow 2018 - Krzysztof Kudryński & Błażej Kubiak (TomTom) - Deep learning in computer vision