Introduction to Neural
Network and Deep
Learning
ISSAM A. AL-ZINATI
OUTREACH & TECHNICAL ADVISOR
UCAS TECHNOLOGY INCUBATOR
ISSAM A. AL-ZINATI - UCASTI 1
ISSAM A. AL-ZINATI - UCASTI 2
Artificial Intelligence vs Machine Learning
Artificial Intelligence is the replication of human intelligence in
computers.
Machine Learning refers to the ability of a machine to learn using
large data sets instead of hard coded rules.
ISSAM A. AL-ZINATI - UCASTI 3
Supervised learning vs unsupervised
learning
Supervised Learning involves using labelled data sets that have
inputs and expected outputs.
Unsupervised Learning is the task of machine learning using data
sets with no specified structure.
ISSAM A. AL-ZINATI - UCASTI 4
What is Deep Learning
ISSAM A. AL-ZINATI - UCASTI 5
Is a Neural Network
What is Deep Learning
ISSAM A. AL-ZINATI - UCASTI 6
Is a Neural Network Neuron
Can run small specific
mathematical task
What is Deep Learning
ISSAM A. AL-ZINATI - UCASTI 7
Is a Neural Network Neuron
Can run small specific
mathematical task
Edge
Connects Neurons
Holds weights to adjust inputs
What is Deep Learning
ISSAM A. AL-ZINATI - UCASTI 8
Is a Neural Network
With More Layers
What is Deep Learning
ISSAM A. AL-ZINATI - UCASTI 9
Is a Neural Network
With More Layers
And More Neurons
What is Deep Learning
Deep Learning is a machine learning method. It allows us to train an
AI to predict outputs, given a set of inputs. Both supervised and
unsupervised learning can be used to train the AI.
ISSAM A. AL-ZINATI - UCASTI 10
How it wok – The Magic
ISSAM A. AL-ZINATI - UCASTI 11
How it work – No Magic
Deep Neural network is not magic. But it is very good at finding patterns.
“The hierarchy of concepts allows the computer to learn complicated concepts
by building them out of simpler ones. If we draw a graph showing how these
concepts are built on top of each other, the graph is deep, with many layers. For
this reason, we call this approach to AI deep learning”, Ian Goodfellow.
Deep Learning is Hierarchical Feature Learning.
ISSAM A. AL-ZINATI - UCASTI 12
How human brain works exactly?
ISSAM A. AL-ZINATI - UCASTI 13
How human brain works exactly?
ISSAM A. AL-ZINATI - UCASTI 14
How perceptron as an artificial neuron
works - Forward neural Network?
ISSAM A. AL-ZINATI - UCASTI 15
How perceptron as an artificial neuron
works - Forward neural Network?
ISSAM A. AL-ZINATI - UCASTI 16
What is weight in Neural Network?
Weight refers to the strength of connection between nodes. Unsigned value
(without +, -) of weight depends on how nodes have power to connect to each
other.
It can be positive or negative. Positive means it is more likely to transmit data
and having strong connection among neurons while negative is vice versa. At the
initialize point we select weight randomly but for having reasonable result it is
better to normalize input data as follow, X is input data:
ISSAM A. AL-ZINATI - UCASTI 17
What is Activation Function role in
Neural Network?
Activation function is (although a bit) equivalent to polarization and stabilizing.
ISSAM A. AL-ZINATI - UCASTI 18
How backward propagation works?
In backward propagation because we need optimum value so we differentiate from sigmoid
function and go inversely from right to left, in order to finding new values for weights.
(1) output_node′ = Sigmoid′ (hidden_sigma) * margin
(2) weight_2 ′ = (output_node′ / hidden_node) + weight_2
(3) hidden_node ′ = (output_node′ / weight_2) * Sigmoid′ (input_sigma)
(4) weight_1 ′ = (hidden_node′ / input_node) + weight_1
(5) Again we repeat steps 1 to 5 with new weights and comparison value from current margin
errors and previous margin errors if current error is less than previous one, so it shows us that
we are in right direction.
(6) We iterate step 1 to 10 until margin error is near to our “Y”.
ISSAM A. AL-ZINATI - UCASTI 19
How backward propagation works?
ISSAM A. AL-ZINATI - UCASTI 20
Why Now
ISSAM A. AL-ZINATI - UCASTI 21
Why Now- Scale
ISSAM A. AL-ZINATI - UCASTI 22
Data
Why Now- Scale
ISSAM A. AL-ZINATI - UCASTI 23
Data
Small Meduim Large
Performance Based on Data Size
Performance
The more data
you feed the
model, the
better results
you get
Why Now- Scale
ISSAM A. AL-ZINATI - UCASTI 24
Model Size & GPU
Why Now- Scale
ISSAM A. AL-ZINATI - UCASTI 25
Model Size & GPU
Small Meduim Large
Performance Based on Model Size
Performance
Bigger model could
achieve better
results.
GPUs help to train
those models in
much faster, 20X!!
Why Now– vs Others
ISSAM A. AL-ZINATI - UCASTI 26
What about other kind of machine learning algorithms, i.e. SVM, DT, Boosting, ….
Would they do better if they got more data and power?
Why Now– vs Others
Small Data Medium Data Large Data
Performance of NN VS Others
Based on Model Size and Data Amount
Others Small NN Medium NN Large NN
ISSAM A. AL-ZINATI - UCASTI 27
Why Now– End-To-End
ISSAM A. AL-ZINATI - UCASTI 28
Usual machine learning approach contains a pipeline of stages that are
responsible of feature extraction.
Each stage passes a set of engineered features which help model to better
understand the case it works on.
This approach is complex and prone to errors.
Why Now– End-To-End
ISSAM A. AL-ZINATI - UCASTI 29
Data (Audio)
Speech Recognition Pipeline
Audio
Features
Phonemes
Language
Model
Transcript
Why Now– End-To-End
ISSAM A. AL-ZINATI - UCASTI 30
Data (Audio)
Speech Recognition - DL
Audio
Features Phonemes Language
Model
Transcript
Why Now– End-To-End
ISSAM A. AL-ZINATI - UCASTI 31
Data (Audio)
Speech Recognition - DL
Audio
Features Phonemes Language
Model
Transcript
The Magic
Deep Learning Models
ISSAM A. AL-ZINATI - UCASTI 32
General
Model
FC
Sequence
Model
RNN
LSTM
Image
Model
CNN
Other
Models
Unsupervised
RL
Deep Learning Models
ISSAM A. AL-ZINATI - UCASTI 33
General
Model
FC
Sequence
Model
RNN
LSTM
Image
Model
CNN
Other
Models
Unsupervised
RL
Hot Research Topic
Advanced Deep Learning Models –
VGGNET - ResNet
Achieves 7.3% on ImageNet-2014 classification Challenge, come in the first
place.
It Used
120 million
parameters.
ISSAM A. AL-ZINATI - UCASTI 34
Advanced Deep Learning Models –
Google Inception V3
Achieves 5.64% on ImageNet-2015 classification Challenge, come in the second place.
ISSAM A. AL-ZINATI - UCASTI 35
Advanced Deep Learning Models –
Google Inception V3
Based on ConvNet concept with the addition
of inception module.
ISSAM A. AL-ZINATI - UCASTI 36
Using a network with a
computational cost of 5 billion
multiply-adds per inference and
with using less than 25 million
parameters.
Deep Learning Applications – Deep Voice
Baidu Research presents Deep Voice, a production-quality text-to-speech system
constructed entirely from deep neural networks.
Ground Truth
Generated Voice
ISSAM A. AL-ZINATI - UCASTI 37
Deep Learning Applications – Image
Captioning
Multimodal Recurrent Neural Architecture generates sentence descriptions from
images. Source.
ISSAM A. AL-ZINATI - UCASTI 38
"man in black shirt is playing guitar." "two young girls are playing with lego toy."
Deep Learning Applications – Generating
Videos
ISSAM A. AL-ZINATI - UCASTI 39
This approach was driven by using Adversarial Network to
1) Generate Videos
2) Conditional Video Generation based on Static Images
Source
Applying Deep Learning – Frameworks
Low Level
ISSAM A. AL-ZINATI - UCASTI 40
Applying Deep Learning – Frameworks
Low Level
ISSAM A. AL-ZINATI - UCASTI 41
High Level
ISSAM A. AL-ZINATI - UCASTI 42
Thanks for listening 

Neural network and deep learning Devfest17

  • 1.
    Introduction to Neural Networkand Deep Learning ISSAM A. AL-ZINATI OUTREACH & TECHNICAL ADVISOR UCAS TECHNOLOGY INCUBATOR ISSAM A. AL-ZINATI - UCASTI 1
  • 2.
  • 3.
    Artificial Intelligence vsMachine Learning Artificial Intelligence is the replication of human intelligence in computers. Machine Learning refers to the ability of a machine to learn using large data sets instead of hard coded rules. ISSAM A. AL-ZINATI - UCASTI 3
  • 4.
    Supervised learning vsunsupervised learning Supervised Learning involves using labelled data sets that have inputs and expected outputs. Unsupervised Learning is the task of machine learning using data sets with no specified structure. ISSAM A. AL-ZINATI - UCASTI 4
  • 5.
    What is DeepLearning ISSAM A. AL-ZINATI - UCASTI 5 Is a Neural Network
  • 6.
    What is DeepLearning ISSAM A. AL-ZINATI - UCASTI 6 Is a Neural Network Neuron Can run small specific mathematical task
  • 7.
    What is DeepLearning ISSAM A. AL-ZINATI - UCASTI 7 Is a Neural Network Neuron Can run small specific mathematical task Edge Connects Neurons Holds weights to adjust inputs
  • 8.
    What is DeepLearning ISSAM A. AL-ZINATI - UCASTI 8 Is a Neural Network With More Layers
  • 9.
    What is DeepLearning ISSAM A. AL-ZINATI - UCASTI 9 Is a Neural Network With More Layers And More Neurons
  • 10.
    What is DeepLearning Deep Learning is a machine learning method. It allows us to train an AI to predict outputs, given a set of inputs. Both supervised and unsupervised learning can be used to train the AI. ISSAM A. AL-ZINATI - UCASTI 10
  • 11.
    How it wok– The Magic ISSAM A. AL-ZINATI - UCASTI 11
  • 12.
    How it work– No Magic Deep Neural network is not magic. But it is very good at finding patterns. “The hierarchy of concepts allows the computer to learn complicated concepts by building them out of simpler ones. If we draw a graph showing how these concepts are built on top of each other, the graph is deep, with many layers. For this reason, we call this approach to AI deep learning”, Ian Goodfellow. Deep Learning is Hierarchical Feature Learning. ISSAM A. AL-ZINATI - UCASTI 12
  • 13.
    How human brainworks exactly? ISSAM A. AL-ZINATI - UCASTI 13
  • 14.
    How human brainworks exactly? ISSAM A. AL-ZINATI - UCASTI 14
  • 15.
    How perceptron asan artificial neuron works - Forward neural Network? ISSAM A. AL-ZINATI - UCASTI 15
  • 16.
    How perceptron asan artificial neuron works - Forward neural Network? ISSAM A. AL-ZINATI - UCASTI 16
  • 17.
    What is weightin Neural Network? Weight refers to the strength of connection between nodes. Unsigned value (without +, -) of weight depends on how nodes have power to connect to each other. It can be positive or negative. Positive means it is more likely to transmit data and having strong connection among neurons while negative is vice versa. At the initialize point we select weight randomly but for having reasonable result it is better to normalize input data as follow, X is input data: ISSAM A. AL-ZINATI - UCASTI 17
  • 18.
    What is ActivationFunction role in Neural Network? Activation function is (although a bit) equivalent to polarization and stabilizing. ISSAM A. AL-ZINATI - UCASTI 18
  • 19.
    How backward propagationworks? In backward propagation because we need optimum value so we differentiate from sigmoid function and go inversely from right to left, in order to finding new values for weights. (1) output_node′ = Sigmoid′ (hidden_sigma) * margin (2) weight_2 ′ = (output_node′ / hidden_node) + weight_2 (3) hidden_node ′ = (output_node′ / weight_2) * Sigmoid′ (input_sigma) (4) weight_1 ′ = (hidden_node′ / input_node) + weight_1 (5) Again we repeat steps 1 to 5 with new weights and comparison value from current margin errors and previous margin errors if current error is less than previous one, so it shows us that we are in right direction. (6) We iterate step 1 to 10 until margin error is near to our “Y”. ISSAM A. AL-ZINATI - UCASTI 19
  • 20.
    How backward propagationworks? ISSAM A. AL-ZINATI - UCASTI 20
  • 21.
    Why Now ISSAM A.AL-ZINATI - UCASTI 21
  • 22.
    Why Now- Scale ISSAMA. AL-ZINATI - UCASTI 22 Data
  • 23.
    Why Now- Scale ISSAMA. AL-ZINATI - UCASTI 23 Data Small Meduim Large Performance Based on Data Size Performance The more data you feed the model, the better results you get
  • 24.
    Why Now- Scale ISSAMA. AL-ZINATI - UCASTI 24 Model Size & GPU
  • 25.
    Why Now- Scale ISSAMA. AL-ZINATI - UCASTI 25 Model Size & GPU Small Meduim Large Performance Based on Model Size Performance Bigger model could achieve better results. GPUs help to train those models in much faster, 20X!!
  • 26.
    Why Now– vsOthers ISSAM A. AL-ZINATI - UCASTI 26 What about other kind of machine learning algorithms, i.e. SVM, DT, Boosting, …. Would they do better if they got more data and power?
  • 27.
    Why Now– vsOthers Small Data Medium Data Large Data Performance of NN VS Others Based on Model Size and Data Amount Others Small NN Medium NN Large NN ISSAM A. AL-ZINATI - UCASTI 27
  • 28.
    Why Now– End-To-End ISSAMA. AL-ZINATI - UCASTI 28 Usual machine learning approach contains a pipeline of stages that are responsible of feature extraction. Each stage passes a set of engineered features which help model to better understand the case it works on. This approach is complex and prone to errors.
  • 29.
    Why Now– End-To-End ISSAMA. AL-ZINATI - UCASTI 29 Data (Audio) Speech Recognition Pipeline Audio Features Phonemes Language Model Transcript
  • 30.
    Why Now– End-To-End ISSAMA. AL-ZINATI - UCASTI 30 Data (Audio) Speech Recognition - DL Audio Features Phonemes Language Model Transcript
  • 31.
    Why Now– End-To-End ISSAMA. AL-ZINATI - UCASTI 31 Data (Audio) Speech Recognition - DL Audio Features Phonemes Language Model Transcript The Magic
  • 32.
    Deep Learning Models ISSAMA. AL-ZINATI - UCASTI 32 General Model FC Sequence Model RNN LSTM Image Model CNN Other Models Unsupervised RL
  • 33.
    Deep Learning Models ISSAMA. AL-ZINATI - UCASTI 33 General Model FC Sequence Model RNN LSTM Image Model CNN Other Models Unsupervised RL Hot Research Topic
  • 34.
    Advanced Deep LearningModels – VGGNET - ResNet Achieves 7.3% on ImageNet-2014 classification Challenge, come in the first place. It Used 120 million parameters. ISSAM A. AL-ZINATI - UCASTI 34
  • 35.
    Advanced Deep LearningModels – Google Inception V3 Achieves 5.64% on ImageNet-2015 classification Challenge, come in the second place. ISSAM A. AL-ZINATI - UCASTI 35
  • 36.
    Advanced Deep LearningModels – Google Inception V3 Based on ConvNet concept with the addition of inception module. ISSAM A. AL-ZINATI - UCASTI 36 Using a network with a computational cost of 5 billion multiply-adds per inference and with using less than 25 million parameters.
  • 37.
    Deep Learning Applications– Deep Voice Baidu Research presents Deep Voice, a production-quality text-to-speech system constructed entirely from deep neural networks. Ground Truth Generated Voice ISSAM A. AL-ZINATI - UCASTI 37
  • 38.
    Deep Learning Applications– Image Captioning Multimodal Recurrent Neural Architecture generates sentence descriptions from images. Source. ISSAM A. AL-ZINATI - UCASTI 38 "man in black shirt is playing guitar." "two young girls are playing with lego toy."
  • 39.
    Deep Learning Applications– Generating Videos ISSAM A. AL-ZINATI - UCASTI 39 This approach was driven by using Adversarial Network to 1) Generate Videos 2) Conditional Video Generation based on Static Images Source
  • 40.
    Applying Deep Learning– Frameworks Low Level ISSAM A. AL-ZINATI - UCASTI 40
  • 41.
    Applying Deep Learning– Frameworks Low Level ISSAM A. AL-ZINATI - UCASTI 41 High Level
  • 42.
    ISSAM A. AL-ZINATI- UCASTI 42 Thanks for listening 