اسلاید روز ششم از کارگاه ۷ روزه دادههای بزرگ و یادگیری ماشین که با تاکید بر یادگیری ژرف برگزار شد. جلسه ششم کارگاه نیز به یادگیری ژرف و کاربردها اختصاص خواهد یافت. این کارگاه به همت ایسیام دانشگاه تهران در محل دانشکده فنی برگزار میشود
زمان هر جلسه ۲ ساعت است
14. 14
ﺑﺎﺯزﻧﻤﺎﯾﯽ ﯾﺎﺩدﮔﯿﺮﯼیﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
The solution is to use machine learning to discover not only the
mapping from representation to output, but also the representation
itself.
15. 15
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
The solution is to use machine learning to discover not only the
mapping from representation to output, but also the representation
itself.
Representation Learning
ﺑﺎﺯزﻧﻤﺎﯾﯽ ﯾﺎﺩدﮔﯿﺮﯼی
16. 16
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
The solution is to use machine learning to discover not only the
mapping from representation to output, but also the representation
itself.
Representation Learning Higher Performance than hand-designed
ﺑﺎﺯزﻧﻤﺎﯾﯽ ﯾﺎﺩدﮔﯿﺮﯼی
Accelerate adaption with
minimal human interaction
17. 17
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
Factors of Variation
When designing features or algorithms for learning features, our goal is
usually to separate the factors of variation that explain the observed data.
18. 18
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
Factors of Variation
Factors are separated sources of influence:
They are not directly observable.
When designing features or algorithms for learning features, our goal is
usually to separate the factors of variation that explain the observed data.
19. 19
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
Factors of Variation
Factors are separated sources of influence:
Such factors are often not quantities.
They are not directly observable.
When designing features or algorithms for learning features, our goal is
usually to separate the factors of variation that explain the observed data.
20. 20
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
Factors of Variation
They may exist either as unobserved objects or unobserved forces in the
physical world that affect observable quantities.
Factors are separated sources of influence:
Such factors are often not quantities.
They are not directly observable.
When designing features or algorithms for learning features, our goal is
usually to separate the factors of variation that explain the observed data.
21. 21
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
Factors of Variation
They may exist either as unobserved objects or unobserved forces in the
physical world that affect observable quantities.
They may also exist as constructs in the human mind that provide useful
simplifying explanations or inferred causes of the observed data.
Factors are separated sources of influence:
Such factors are often not quantities.
They are not directly observable.
When designing features or algorithms for learning features, our goal is
usually to separate the factors of variation that explain the observed data.
22. 22
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
Factors of Variation
Factors of Variation:
Position
23. 23
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
Factors of Variation
Factors of Variation:
Position
Color/texture
24. 24
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
Factors of Variation
Factors of Variation:
Position
Color/texture
The angle/brightness of the sun
25. 25
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
Factors of Variation
Factors of Variation:
Position
Color/texture
The angle/brightness of the sun
The weather (Snow/Rain/Fog/Dust)
26. 26
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
Factors of Variation
Factors of Variation:
Position
Color/texture
The angle/brightness of the sun
The weather (Snow/Rain/Fog/Dust)
The amount of load
27. 27
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
Factors of Variation
Factors of Variation:
Position
Color/texture
The angle/brightness of the sun
The weather (Snow/Rain/Fog/Dust)
The amount of load
Exterior Accessories
28. 28
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
Factors of Variation
Factors of Variation:
Position
Color/texture
The angle/brightness of the sun
The weather (Snow/Rain/Fog/Dust)
The amount of load
Exterior Accessories
The individual pixels in an
image of a red car might be
very close to black at night…
29. 29
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
Factors of Variation:
Position
Color/texture
The angle/brightness of the sun
The weather (Snow/Rain/Fog/Dust)
The amount of load
Exterior Accessories
The individual pixels in an
image of a red car might be
very close to black at night…
ﮐﺮﺩد؟ ﺑﺎﯾﺪ ﭼﻪ
35. 35
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
ﺑﺎﺯزﻧﻤﺎﯾﯽ + ﺳﺎﺯزﯼیﻩهﺳﺎﺩد ﺑﺮﺍاﯼی ﺗﮑﻨﯿﮏ ﯾﮏ
Raw Sensory Input Data
Car
Person
Animal
36. 36
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
ﺑﺎﺯزﻧﻤﺎﯾﯽ + ﺳﺎﺯزﯼیﻩهﺳﺎﺩد ﺑﺮﺍاﯼی ﺗﮑﻨﯿﮏ ﯾﮏ
Raw Sensory Input Data The function mapping from a set of pixels
to an object identity is very complicated.
Car
Person
Animal
!ﻧﯿﺴﺖ ﻣﻤﮑﻦ ﻣﺴﺘﻘﯿﻢ ﺗﺒﺪﯾﻞ
48. 48
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
.ﺍاﺳﺖ Autoencoder ،ﺑﺎﺯزﻧﻤﺎﯾﯽ ﯼیﯾﺎﺩدﮔﯿﺮ ﺍاﻟﮕﻮﺭرﯾﺘﻢ ﯾﮏ ﺟﻮﻫﺮ
Autoencoder
Input = decoder(encoder(input))
An autoencoder is the combination of an encoder function that converts
the input data into a different representation, and a decoder function that
converts the new representation back into the original format.
49. 49
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
.ﺍاﺳﺖ Autoencoder ،ﺑﺎﺯزﻧﻤﺎﯾﯽ ﯼیﯾﺎﺩدﮔﯿﺮ ﺍاﻟﮕﻮﺭرﯾﺘﻢ ﯾﮏ ﺟﻮﻫﺮ
Autoencoder
Input = decoder(encoder(input))
Reconstructed Original
An autoencoder is the combination of an encoder function that converts
the input data into a different representation, and a decoder function that
converts the new representation back into the original format.
56. 56
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
ﺍاﺑﺰﺍاﺭر ﯾﮏ ﺑﺎ ﺁآﺷﻨﺎﯾﯽ - ﻣﺎﺷﯿﻦ ﯾﺎﺩدﮔﯿﺮﯼی
What is Skytree Express?
• Skytree Express is an express way to use Skytree’s machine learning
software. It is available for use via a Virtual Machine (VM) for Windows,
MAC OS X, and through an easy install script on RHEL/CentOS systems.
• Currently, Skytree Express comes in two versions (i) Skytree GUI & Python
SDK and (ii) Skytree Command Line Interface (CLI).
• Skytree Express can be downloaded free of cost. It comes with a
preconfigured license that is valid for one year.
64. 64
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
ﺍاﺑﺰﺍاﺭر ﯾﮏ ﺑﺎ ﺁآﺷﻨﺎﯾﯽ - ﻣﺎﺷﯿﻦ ﯾﺎﺩدﮔﯿﺮﯼی
Skytree Express is available for personal, educational, and even
commercial usage! The free version restricts usage up to 100 million
elements on a single machine/node.
License:
87. 87
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
ﺑﺪﻫﯿﻢ ﯾﺎﺩد ﺧﻮﺍاﻧﺪﻥن ﻣﺎﺷﯿﻦ ﺑﻪ
To feed an image into our neural network, we simply treat
the 18x18 pixel image as an array of 324 numbers:
88. 88
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
ﺑﺪﻫﯿﻢ ﯾﺎﺩد ﺧﻮﺍاﻧﺪﻥن ﻣﺎﺷﯿﻦ ﺑﻪ
The handle 324 inputs, we’ll just enlarge our neural network to have 324 input nodes:
89. 89
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
ﺑﺪﻫﯿﻢ ﯾﺎﺩد ﺧﻮﺍاﻧﺪﻥن ﻣﺎﺷﯿﻦ ﺑﻪ
All that’s left is to train the neural network with images of “8”s and not-“8"s so it learns to tell
them apart. When we feed in an “8”, we’ll tell it the probability the image is an “8” is 100% and
the probability it’s not an “8” is 0%.
Some training data:
110. 110
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
ﺑﺪﻫﯿﻢ ﯾﺎﺩد ﺧﻮﺍاﻧﺪﻥن ﻣﺎﺷﯿﻦ ﺑﻪ
As a human, you intuitively know that pictures have a hierarchy or conceptual structure.
111. 111
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
ﺑﺪﻫﯿﻢ ﯾﺎﺩد ﺧﻮﺍاﻧﺪﻥن ﻣﺎﺷﯿﻦ ﺑﻪ
As a human, you intuitively know that pictures have a hierarchy or conceptual structure.
As a human, you instantly recognize the hierarchy in this picture:
• The ground is covered in grass and concrete
• There is a child
• The child is sitting on a bouncy horse
• The bouncy horse is on top of the grass
112. 112
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
ﺑﺪﻫﯿﻢ ﯾﺎﺩد ﺧﻮﺍاﻧﺪﻥن ﻣﺎﺷﯿﻦ ﺑﻪ
As a human, you intuitively know that pictures have a hierarchy or conceptual structure.
Mad scientists literally poking cat brains with weird probes
to figure out how cats process images
121. 121
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
ﺑﯿﺸﺘﺮ ﻫﺎﯼیﻪﻻﯾ ﺍاﻓﺰﻭوﺩدﻥن
Our image processing pipeline is a series of steps:
• convolution,
• max-pooling,
• a fully-connected network.
122. 122
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
ﺑﯿﺸﺘﺮ ﻫﺎﯼیﻪﻻﯾ ﺍاﻓﺰﻭوﺩدﻥن
The first convolution step might learn to recognize sharp edges, the
second convolution step might recognize beaks using it’s knowledge
of sharp edges, the third step might recognize entire birds using it’s
knowledge of beaks, etc.
The more convolution steps you have, the more complicated
features your network will be able to learn to recognize.
123. 123
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
ﺑﯿﺸﺘﺮ ﻫﺎﯼیﻪﻻﯾ ﺍاﻓﺰﻭوﺩدﻥن
In this case, they start a 224 x 224 pixel image, apply convolution and max pooling
twice, apply convolution 3 more times, apply max pooling and then have two fully-
connected layers. The end result is that the image is classified into one of 1000
categories!
124. 124
ﺍاﻥنﺮﺗﻬ ﺩدﺍاﻧﺸﮕﺎﻩه ACM - ۱۳۹۵ ﺗﺎﺑﺴﺘﺎﻥن - ﻣﺎﺷﯿﻦﯼیﯾﺎﺩدﮔﯿﺮ ﻭو ﺑﺰﺭرﮒگ ﻫﺎﯼیﻩهﺩدﺍاﺩد ﻫﺎﯼیﺩدﮐﺎﺭرﺑﺮ ﺑﺮ ﮔﺬﺭرﯼی
ﮐﻨﯿﻢ؟ ﺷﺮﻭوﻉع ﮐﺠﺎ ﺍاﺯز
So how do you know which steps you need to combine to make
your image classifier work?
Honestly, you have to answer this by doing a lot of experimentation and
testing. You might have to train 100 networks before you find the optimal
structure and parameters for the problem you are solving. Machine
learning involves a lot of trial and error!