O slideshow foi denunciado.
Utilizamos seu perfil e dados de atividades no LinkedIn para personalizar e exibir anúncios mais relevantes. Altere suas preferências de anúncios quando desejar.
Louis Monier
@louis_monier
https://www.linkedin.com/in/louismonier
Deep Learning for Images
I see what you mean...
Gregory...
Fun with Images
Image Classification: kayak, boy Entity Detection
kayak boy
Face Recognition
Leo
Gollum
More Fun with Images
Pose DetectionImage SegmentationImage Captioning: “A young boy
wearing an orange vest riding a
yellow...
Yet more Fun with Images
Optical Character Recognition (OCR):
Astronomy is the science which treats of the nature and
prop...
Our Wet Hardware
Alternating layers of
- simple cells (filters)
- complex cells (combination)
Simple patterns to abstract ...
Convolutional Neural Network (ConvNet, CNN)
Suggested by Kunihiko Fukushima, 1980
LeNet, by Yann LeCun, 1998, to classify ...
filterimage
= 6.6
= -7.8
1.0 - really want
0.2 - sort of want
-1.0 - don’t want
Convolution: Applying a Filter to a Signal...
Convolutional Layer - Basic Unit
5x5x3 chunk of inputs
Layer N Layer N+1
ReLU
neuron
(3)
(3)
(3)
(3)
5 x 5 x 3 = 75 inputs...
Convolutional Layer: Add Depth
5x5x3 chunk of inputs
Layer N Layer N+1
Depth = 7 ReLU neurons in parallel,
with different ...
Stride = 1
Convolutional Layer: Repeat over entire image
L=W=5
D=4
zero
padding
D=7
Shared weights!!!
Pooling Layer: Squeeeeeze!
Max Pooling Average Pooling
Layer N+1Layer N
Classical CNN topology - VGGNet (2013)
224x224 112x112 56x56 28x28 14x14 FC
D=64
D=128
D=256
D=512
D=512
D=4096 D=4096 D=1...
Layer 1
Filter Matching images
Layer 2
Layer 3
Layer 4
Layer 5
Modern ConvNet - GoogLeNet
GoogLeNet (2014)
ResNet-34 (2015)
Manifolds
Real-life Data vs Random Data
If music be the food of love, play on!
-- William Shakespeare
3Flr'kI5;LS3oLj1xK52,BA1 Rea5I...
Workshop : Keras & MNIST
https://github.com/holbertonschool/deep-learning/tree/master/Class%20%232
Workshop : Keras & CIFAR 10
https://github.com/holbertonschool/deep-learning/tree/master/Class%20%232
Deep Learning Class #2 - Deep learning for Images, I See What You Mean
Próximos SlideShares
Carregando em…5
×

Deep Learning Class #2 - Deep learning for Images, I See What You Mean

Slides by Louis Monier (Altavista Co-Founder & CTO) for Deep Learning keynote #2 at Holberton School. http://www.meetup.com/Holberton-School/events/230547621/

The keynote was followed by a workshop prepared by Gregory Renard. If you want to assist to similar keynote for free, checkout http://www.meetup.com/Holberton-School/

  • Entre para ver os comentários

Deep Learning Class #2 - Deep learning for Images, I See What You Mean

  1. 1. Louis Monier @louis_monier https://www.linkedin.com/in/louismonier Deep Learning for Images I see what you mean... Gregory Renard @redo https://www.linkedin.com/in/gregoryrenard Class 2 - Q2 - 2016
  2. 2. Fun with Images Image Classification: kayak, boy Entity Detection kayak boy Face Recognition Leo Gollum
  3. 3. More Fun with Images Pose DetectionImage SegmentationImage Captioning: “A young boy wearing an orange vest riding a yellow kayak on water, with sunlight reflections.”
  4. 4. Yet more Fun with Images Optical Character Recognition (OCR): Astronomy is the science which treats of the nature and properties of the heavenly bodies. Autonomous Vehicles Handwriting Recognition: combustible: “able to catch fire”, adjective for being capable of igniting and burning.
  5. 5. Our Wet Hardware Alternating layers of - simple cells (filters) - complex cells (combination) Simple patterns to abstract concepts ~ 5B neurons for vision
  6. 6. Convolutional Neural Network (ConvNet, CNN) Suggested by Kunihiko Fukushima, 1980 LeNet, by Yann LeCun, 1998, to classify hand-written digits
  7. 7. filterimage = 6.6 = -7.8 1.0 - really want 0.2 - sort of want -1.0 - don’t want Convolution: Applying a Filter to a Signal through through = = image filter 1.0 0.5 0.0 through through
  8. 8. Convolutional Layer - Basic Unit 5x5x3 chunk of inputs Layer N Layer N+1 ReLU neuron (3) (3) (3) (3) 5 x 5 x 3 = 75 inputs 76 weights
  9. 9. Convolutional Layer: Add Depth 5x5x3 chunk of inputs Layer N Layer N+1 Depth = 7 ReLU neurons in parallel, with different weights
  10. 10. Stride = 1 Convolutional Layer: Repeat over entire image L=W=5 D=4 zero padding D=7 Shared weights!!!
  11. 11. Pooling Layer: Squeeeeeze! Max Pooling Average Pooling Layer N+1Layer N
  12. 12. Classical CNN topology - VGGNet (2013) 224x224 112x112 56x56 28x28 14x14 FC D=64 D=128 D=256 D=512 D=512 D=4096 D=4096 D=1000 FC FC + Softmax ConvNet Pool
  13. 13. Layer 1 Filter Matching images
  14. 14. Layer 2
  15. 15. Layer 3
  16. 16. Layer 4
  17. 17. Layer 5
  18. 18. Modern ConvNet - GoogLeNet GoogLeNet (2014) ResNet-34 (2015)
  19. 19. Manifolds
  20. 20. Real-life Data vs Random Data If music be the food of love, play on! -- William Shakespeare 3Flr'kI5;LS3oLj1xK52,BA1 Rea5IYSf -- 1000 monkeys typing -- Real world -- Random Pixels
  21. 21. Workshop : Keras & MNIST https://github.com/holbertonschool/deep-learning/tree/master/Class%20%232
  22. 22. Workshop : Keras & CIFAR 10 https://github.com/holbertonschool/deep-learning/tree/master/Class%20%232

×