Artificial Intelligence (AI) & Fifth Generation (5G)...
Transcript of Artificial Intelligence (AI) & Fifth Generation (5G)...
Artificial Intelligence (AI) & Fifth Generation (5G) Networks
Choong Seon Hong
Department of Computer Science and EngineeringKyung Hee University
• Introduction• Background
• 5G deliverables
• AI deliverables
• Motivation of AI in 5G
• Artificial Intelligence (AI) • Introduction
• Evolution of AI
• Machine Learning
• Artificial Neural Networks
• Deep Learning
• Deep Learning implementation using open source: Use-Cases
• Applications of AI in 5G• Network Slicing Enablers
• Wireless Network Virtualization
• Evolution of Cellular Networks
• Network slicing Deliverables
• Network Slicing Industrial Efforts
• AI and Network Slicing for 5G Networks: Use-Cases
• Conclusions
Outline 2
Introduction• Background• 5G deliverables• AI deliverables• Motivation of AI in 5G
3
Background
• Traffic growth due to:• Tsunami of heterogeneous connections:
• Smartphones
• Connected vehicles
• Wearables devices
• IoT sensors
• And so on…
• Novel bandwidth hungry applications:
• Real time HD streaming
• Online Gaming
• Ultra-reliable and low-latency
communication
• Virtual reality services
• Enhanced mobile broadband
• And so on…
Source: Cisco Visual Networking Index (VNI), Feb. 2017.
4
Bottleneck and New Paradigms in 5G
• Network capacity is a
bottleneck due to:• Radio Access Networks (RANs)
• Mostly wireless and highly dynamic
• New paradigms to support:• Small cell (SC) deployment
• Device-to-device(D2D)
• Network virtualization
• LTE-unlicenced
• And so on…
.
Source: http://www.eurescom.eu/news-and- events/eurescommessage/eurescom-message-1-2014/3gpp-system-standards-heading-into-the-5g-era.html.
5
Introduction: 5G?
Broadband Back bone
6
A. Agarwal, G. Misra and K. Agarwal, "The 5th Generation Mobile Wireless Networks- Key Concepts, Network Architecture and Challenges", American Journal of Electrical and Electronic Engineering, Vol.3, No.2, pp.22-28, 2015.
5G Deliverables
5G Deliverables: • Higher data rates
• Reduced end-to-end latency
• Higher energy efficiency
• Better network coverage
• Enhanced security
• Ultra reliability
• and so on…
Source: “5G Use Cases and Requirements,” a white paper from Nokia.
7
• KT, SK-Telecom, LGU+, Korea:
• Successfully collaborated with Samsung Electronics and Huawei to commertialize a 5G end-to-end network that includes:
• 5G virtualized core
• Virtualized RAN
• Distributed Unit
(baseband unit and radio unit)
• Devices - based on the 3GPP
5G New Radio (5G NR)
5G Networks: Industrial efforts 8
• Is it 20G? (LTE 1G)
Current Commencialized Speed 9
• Nokia, Finland:
• Actively focusing for providing 5G services such as:
• 5G mobility service supporting enhanced mobile broadband (eMBB)
• 5G mobility service supporting ultra-reliable and ultra-low latency communications (URLLC)
• Huawei, China:
• Huawei have actively worked to enhance the antenna capabilities for 5G networks
• It has released its new FDD antenna and FDD/TDD converged antenna platforms
• And it developed Switching systems and Base Station
5G Networks: Industrial efforts 10
• Ericsson, Sweden:
• Ericsson’s is actively participating in the development of 5G networks
• It has recently developed the Ericsson’s 5G radio test-bed that comprises of Massive MIMO, multi-user MIMO, and beamforming technologies
• Ericsson has introduced a new radio product, AIR 3246, for Massive Multiple Input Multiple Output (Massive MIMO).
• This launch will enable operators – especially in metropolitan areas – to bring 5G to subscribers using today’s mid-band spectrum and boost capacity in their LTE networks.
5G Networks: Industrial efforts 11
• Air Latency 1ms
• Transmission Latency << 1ms
• Core Latency 1ms
• App./Server E2E Latency 10ms
Low Latency in 5G 12
• Total E2E Latency 13ms
• How low latency is needed?
Massive Connectivity 13
Multi-RAT Aggregation 14
• Spectrum : 28GHZ
5G of Verizon 15
Introduction: Artificial Intelligence
• Artificial Intelligence (AI) is the science and engineering of making
machines as intelligent as humans.
• AI Deliverables:• Create systems that can perform:
• Perception.
• Reasoning.
• Learning.
• Problem Solving.
16
• To unleash the true potential of 5G networks:
• Intelligent functions using AI across both the edge and core of the network are required along with the novel enabling technologies.
• AI functions must be able to:
• Adaptively exploit the wireless system resources.
• Generated data to optimize network operation.
• Guarantee the QoS in real time.
• Mobile edge and core intelligence can only be realized by integrating fundamental notions of artificial intelligence (AI) across the wireless infrastructure and end-user devices.
Introduction: Motivation of AI in 5G 17
• Role of AI in 5G networks: Exploit big data analytics to enhance situational awareness and overall network operation such as:
• Fault Monitoring
• User Tracking
• Cell Association
• Radio Resource Management
• Cache Resource Management
• Mobility Management
• Management and Orchestration
• Service Provisioning Management
• And so on..
Introduction: Motivation of AI in 5G 18
Artificial Intelligence (AI) • Introduction• Evolution of AI• Machine Learning • Artificial Neural Networks• Deep Learning• Deep Learning implementation using open source: Use-Cases
19
Artificial Intelligence, Machine Learning and Deep Learning 20
http://www.deeplearningbook.org/contents/intro.html
Machine Learning
AI
Deep Learning
AI – Any technique which enables computers to mimic human behavior.
ML – Subset of AI techniques which use statistical methods to enable machines to improve with experiences.
DL – Subset of ML which make the computation of multi-layer neural networks feasible.
• In the early days of AI, researchers were very interested in machines that could learn from data.
• But ML’s increased focus on a logical, knowledge-based approach occasioned a split from AI in 1980.
• Statistical-based research, probabilistic reasoning, pattern recognition and information retrieval came into the fold of ML; by 1990s, ML became a separate field altogether, and began to flourish by shifting its goal from achieving AI to tackling practical problems.
• Machine Learning grew out a branch of artificial intelligence that studies pattern recognition and computational learning. It is a subfield of computer science.
History 21
The Evolution of Artificial Intelligence (AI) 22
1950Alan Turing
Proposes the
Turing Test
1950Isaac Asimov
proposes the Three
Laws of Robotics
1951First AI based
Program was
written
1955First self learning
game playing
program is written
1959MIT AI Lab
is setup
1961First Robot is inducted
into GM’s assembly
production line
1963First Machine
Learning program
is written
1964First demonstration of
an AI program which
understand Natural
Language
1965First AI based
Chat-bot
(ELLZA) was
created
1969Stanford Research
Institute (SRI)
demonstrates the first
locomotive and intelligent
robot (Shakey)
1969First autonomous
vehicle is created at the
Stanford AI LAB
1974First rule based AI
expert system for
medical diagnostics
1980LISP based machines
are developed and
marketed
1986Learning
representations by
back-propagating error
(Backpropagation)
1997IBMs Deep Blue beats
Gary Kasparov at
Chess
1999Sony introduces the
first artificially
intelligent domestic
robot, AIBO
1999First Emotional AI
machines
demonstrated at
MIT AI Lab
2004DAPRA introduces
the first challenge
for Autonomous
Vehicles
2005AI based
recommendation
engines
2009Google builds
Self Driving Car
2010Narrative Science’s
AI demonstrates
ability to write
reports
2011IBM Watson beats
Jeopardy
champions
2011Personal
Assistants like Siri,
Google Now and
Cortana become
mainstream
2015Elon Musk and others
announce a $1B non
profit open source
initiative, OPEN AI to
develop friendly AI
2016Google’s
Deepmind AlphaGo
defeats Go’s
champions
2016NVIDIA announces
supercomputer for
Deep Learning and
AI
Modified from Source: https://twitter.com/mikequindazzi/status/835589969909424130
2017AlphaGo Zero which
learns from scratch
2018OpenAI 5
(Dota 2)
AI Revenue, Top 10 Use Cases, World Markets, 2016-2025 23
Source :https://twitter.com/MikeQuindazzi
Reasoning/Prediction: It is the set of processes that enables us to provide basis for making decisions, and prediction.
Problem Solving: It is the process in which one perceives and tries to arrive at a desired solution from a current situation.
Perception: The process of acquiring, interpreting, selecting, and organizing sensory information.
Learning: It is the process of knowledge acquisition by experiencing.
Overview of AI architecture 24
Artificial Intelligence
Reasoning
Problem Solving
Perception
Learning
• Perception: Perception presumes sensing. In humans, perception is aided
by sensory organs.
• In the domain of AI, perception mechanism puts the data acquired by the artificial sensors together to analyze the environment into objects and their features and relationships.
• Example
• Artificial perception is sufficiently well advanced to enable optical sensors to identify individuals such as autonomous vehicles to drive at moderate speeds on the open road, and robots to roam through buildings collecting trash.
Components of Artificial Intelligence: Perception 25
• Learning: It is the activity of gaining knowledge or skill by studying, practicing, being taught, or experiencing something.
• The simplest learning is by trial-and-error method.
• Learning is categorized as
• Supervised learning
• Unsupervised learning
• Reinforcement learning
• Example
• A simple program for solving mate-in-one chess problems might try out moves at random until one is found that achieves mate.
Components of Artificial Intelligence: Learning 26
• Reasoning/Prediction: It is the set of processes that enables us to provide basis for making decisions, and prediction.
• To reason is to draw inferences appropriate to the situation.
• Deductive Reasoning
• Inductive Reasoning
• Case-based reasoning
• Rule-based reasoning
• Example:
• Previous accidents of these types were caused by instrument failure; therefore, we can reason and predict that this accident was also caused by instrument failure.
Components of Artificial Intelligence: Reasoning 27
• Problem Solving: It is the process in which one perceives and tries to arrive at a desired solution from a current situation.
• Problem solving also includes decision making, which is the process of selecting the best suitable alternative out of multiple alternatives to reach the desired goal are available.
• Some problem solving AI techniques are:
• Genetic algorithms
• Fuzzy models
• Swarm intelligence
• Example:
• Some examples are finding the winning move (or sequence of moves) in a board game, devising mathematical proofs, and manipulating “virtual objects” in a computer-generated world.
Components of Artificial Intelligence: Problem Solving 28
• Machine learning (ML) deals with systems and algorithms that can learn from various data and make predictions.
• Example: Predicting traffic patterns at a busy intersection
• A program can run a machine learning algorithm containing past data with traffic patterns and, having “learned” previous data, it can devise better predictions of future traffic patterns
Machine Learning (ML) 29
https://futurism.com/images/understanding-machine-learning-infographic/
• The importance of ML is that, since it’s data-driven, it can be trained to create valuable predictive models that can guide proper decision and smart actions
• With today’s processing, and cheaper data storage, it’s possible to build models that can analyze large, complex data to deliver faster and more precise results
Applications 30
https://futurism.com/images/understanding-machine-learning-infographic/
Machine Learning Taxonomy 31
Machine Learning Types
No target variableCategorical
target variable
Continuoustarget variable
Categorical target variable
No target variable
LogisticLinear
Regression
Housing price prediction
SVM kNN
Classification
Medical Imaging
K-means Hierarchical
Clustering
Customer Segmentation
AprioriDBSCAN
Association
Market Basket Analysis
Hybrid RL+SVM RL+NN
Classification
Optimized Marketing
Deep Q-learning Actor Critic learning
Control
Self-Driving Cars
K-means, HMM, CRF, MEMM, GMM
Unsupervised learning
MDP, Markov approximation, Q-learning
Reinforcement learning
SVM, kNN, Naïve Bayes, Random Forest
Supervised learning
• Supervised ML – relies on data where the true label is indicated. Example: teaching a computer to distinguish between pictures of cats and dogs, with each image tagged “cat” or “dog”. Labeling is normally performed by humans to guarantee high data quality. Having learned the difference, the ML algorithm can now classify new data and predict labels (“cat” or “dog”) on previously unseen images.
Supervised ML 32
• Supervised learning:
• Supervised learning algorithms are trained using labeled data.
• When dealing with labeled data, both the input data and its desired output data are known to the system.
• Supervised learning is commonly used in applications that have enough historical data.
• Applications:
• Classification
• Regression
Machine Learning: Supervised learning
Machine Learning
Algorithm
Predictive Model
features vectorNew
Text, Document,
Image, Sound
Training Text,
Documents, Images,
Sounds…
features vector
33
Expected Label
Labels
Machine Learning Model (Supervised) 34
Training
Model
Output
Evaluation
Features
Extraction
Weight update
Raw
DATA
Labels
New
DATA
ModelFeatures
ExtractionOutput
Predicted
Outputs
Save Train Model
Training phase
Testing phase
• Model Selection(in Training Model) - Convolutional Neural Network, Recurrent Neural
Network, etc..
• Performance Metrics – Accuracy
• Feature Extraction :Scaling (normalized
inputs) (Dimensionality Reduction
• Feature Selection : Selecting important
features
• Algorithms: • Decision tree
• Random forest
• Neural networks
• Support vector machines
• Ensemble learning
• Bayesian learning
• And so on..
• Examples:
• Speech recognition used in smart devices is based on supervised learning techniques.
Machine Learning: Supervised learning 35
• Supervised Learning is being used to detect spam emails, i.e., Naive Bayes spam filtering. Particular words have specific probabilities of occurring in spam email and in legitimate email, e.g., “refinance”, “Viagra”
• Probabilities are not known in advance. A filter is trained by users manually indicating if email is spam or not through which the filter adjusts the probabilities of each word and save in its database
• After training, the word probabilities are used to compute the probability that an email with a particular set of words belongs to either spam or not spam category
Example: Supervised learning for spam filtering 36
!!!!$$$!!!!
Spam filtering
Input Output
Spam
or
Not Spam
• Unsupervised ML- deprives a learning algorithm of the labels used in supervised learning. Usually involves providing the ML algorithm with a large amount of data on every aspect of an object. Example: presented with images of cats and dogs that have not been labeled, unsupervised ML can separate the images into two groups based on some inherent characteristics of the images.
Unsupervised ML 37
Machine Learning: Unsupervised learning
• Unsupervised learning:
• Unsupervised learning is a type of machine learning algorithm used to draw inferences from datasets consisting of input data.
• No labels are given to the learning algorithm, leaving it on its own to find structure in its input.
• Unsupervised learning can be a goal in itself to discover hidden patterns in data.
• Applications:
• Clustering
• Associations
• Anomaly detection
38
Machine Learning: Unsupervised learning
• Algorithms: • Kmeans clustering• Hierarchical clustering• DBScans• Apriori Associations• Principal component analysis• Independent component analysis • Non-negative matrix factorization• And so on..
• Examples:• Market segmentation uses clustering to identify subgroups of people who might be
more receptive to a specific form of advertising, or more likely to purchase a particular product.
• In medicine, clustering diseases, cures for diseases, or symptoms of diseases can lead to very useful taxonomies.
39
Machine Learning: Reinforcement learning
• Reinforcement learning:
• Reinforcement learning is a type of learning in which an agent learns its best action through trial-and-error by interactions with a dynamic environment.
• Reinforcement Learning is learning how to act in order to maximize a numerical reward.
• Close to human learning.
• Every action has some impact in the environment, and the environment provides rewards that guides the learning algorithm.
• Applications:
• Delivery Management
• Supply chain inventory management
• Stock market trading
40
• Reinforcement Learning – Example: learning to play chess. ML receives information about whether a game played was won or lost. The program does not have every move in the game tagged as successful or not, but only knows the result of the whole game. The ML algorithm can then play a number of games, each time giving importance to those moves that result in a winning combination.
Reinforcement Learning 41
Machine Learning: Reinforcement learning
• Algorithms: • Q-Learning
• Double Q-Learning
• Actor critic learning
• State–action–reward–state–action (SARSA)
• Expected SARSA
• Temporal-Difference Learning
• And so on..
• Examples:• A robot uses deep reinforcement learning to pick a device from one box and
putting it in a container. Whether it succeeds or fails, it memorizes the object and gains knowledge and train’s itself to do this job with great speed and precision.
42
• Artificial Neural Networks – a learning algorithm, inspired by biological neurons, that uses statistical data modeling tools to find patterns in data.
• ANNs are capable of applying the human intelligence to learn complex patterns and to model relationships that are too difficult to learn by traditional techniques
• The key technique of ANNs is multiple processing elements called neurons working in parallel to solve a specific problem
Artificial Neural Networks (ANN)
x1
x2
x3
xn
Activation Function
Output
.
.
.
43
• The perceptron is a mathematical model of a biological neuron.
Perceptron 45
https://cs.stanford.edu/people/eroberts/courses/soco/projects/neural-networks/Neuron/index.html
Output
w1
w2
w3
x1
x2
x3
Activation Function Equation Example Graph
Unit Step
𝜙 𝑧 = ቐ0, 𝑧 < 00.5, 𝑧 = 01, 𝑧 > 0
Perceptron
Linear 𝜙 𝑧 = zAdaline(adaptive linear neuron), Linear regression
Sigmoid𝜙 𝑧 =
1
1 + 𝑒−𝑧Perceptron,Logistic Regression,Multi layer Neural Network
Rectified Linear Unit (ReLU) 𝜙 𝑧 = max(0, z) Perceptron,Multi layer Neural Network
Examples of Activation functions in Perceptron 46
Perceptron and ANN 47
The main components of a ANN are as follows • Input Layer• Output Layer• Hidden Layer• Input Weight Matrix• Output Weight Matrix
A neuron or perceptron A neural network
mainly consists of Input/output and an activation function
consists of connecting these neurons through multiple weights .
Main types of Artificial Neural Networks are
• Feed Forward Neural Networks:
Connections between neurons come only from the previous layer with no feedback involved.
• Recurrent Neural Networks:
Connection between neurons forms a directed cycle so as to use the internal memory of each unit for information processing.
Artificial Neural Networks (ANN): Classifications 48
Input layer (x) Hidden layer (h)Output layer (y)
Feed-forward networks have the following characteristics:
1. Perceptrons are arranged in layers, with the first layer taking in inputs and the last layer producing outputs. The middle layers have no connection with the external world, and hence are called hidden layers.
2. Each perceptron in one layer is connected to every perceptron on the next layer. There is no connection among perceptrons in the same layer.
3. Drawback of FNN is that it is not capable of handling the order of input samples, i.e., sequence of time.
Feed Forward Neural Network (FNN)
Input HiddenOutput
49
Recurrent Neural Networks (RNNs) 50
RNNs have the following characteristics:
1. RNNs, are designed for modeling sequences and are capable of remembering past information and processing new events accordingly.
2. In RNNs, the hidden layer gets its input from both the input layer and the hidden layer from the previous time step.
3. RNNs have multiple categories:
• One to many model
• Many to one model
• Many to many model
• Learning requires adjustment of weights between neurons which is called training
• The objective is to minimize the error (𝐸) between desired and actual output which is:
𝐸 𝑾, 𝒃 =1
2(||𝒚 𝑾, 𝒃, 𝒙 − 𝒚𝐷||)
• Gradient descent is commonly used to minimize 𝐸 𝑾,𝒃
Training in ANNs 51
𝑾 is weight matrix
𝒚 𝑾, 𝒃, 𝒙 is desired output
𝒚𝐷 is the actual output
𝒙 is input vector𝒃 is bias vector
𝜸 is learning rate
represents the partial derivative
Shallow and Deep ANNs 52
𝑥1
𝑥2
𝑥3
ℎ1
ℎ2
ℎ3
ℎ4
ො𝑦1
ො𝑦2
Input LayerHidden Layer
Output Layer
Shallow Artificial Neural Network
𝑤11
𝑤14
𝑤21
𝑤42
𝑥1
𝑥1
𝑥1
ℎ11
ℎ12
ℎ13
ℎ14
ℎ21
ℎ22
ො𝑦1
ො𝑦2
ො𝑦3
Input Layer Hidden Layers Output Layer
Deep Artificial Neural Network
𝑤1,11
𝑤1,14
𝑤11,21
𝑤14,22
𝑤21,1
𝑤22,3
• ANNs with one hidden layer are typically called shallow neural networks
• Deep Neural Networks (DNNs) have many hidden layers for learning multiple levels of representation and abstraction
• Initially DNN could not be used to model the problems due to high number of parameters and its computation in real-time.
• Deep learning can now be realized because of the followings:
• Improved computing capacity
• Improved datasets, i.e., Big Data
• Improved training algorithms and network architectures
Motivation for Deep Neural Network 53
Milestones in the Development of Deep Neural Networks 54
https://beamandrew.github.io/deeplearning/2017/02/23/deep_learning_101_part1.htmlhttps://deepmind.com/research/alphago/
Geoffrey Hinton
Deep Learning Taxonomy 55
Modified from source: Fadlullah, Zubair, et al. "State-of-the-Art Deep Learning: Evolving Machine Intelligence Toward Tomorrow’s Intelligent Network Traffic Control Systems." IEEE Communications Surveys & Tutorials (2017).
Deep Learning
Reinforcement learning
Unsupervised Learning
Supervised learning
• Deep Q-learning• Double Q-learning• Prioritized experience replay
Reinforcement Learning
Advertising and business intelligence (Google ads, etc.), Weather forecasting, Market forecasting, Political campaigns
Real-time decisions, Game Artificial Intelligence, Learning tasks, Skill acquisition, Personal assistants (Google Now, Microsoft Cortana, Apple Siri, etc.), Autonomous (“Self-driving”) cars
Big data visualization, Feature elicitation, Structure discovery, Meaningful compression
Recommendation engines (Amazon web service, Netflix, etc.), Customer segmentation, Target marketing, Filter
Economics (risk prediction, etc.)
• Neural Network (NN)• Convolutional Neural Network (CNN)• Deep Belief Networks (DBN)• Recurrent Neural Network (RNN)
Classification
• Neural Network (NN)
Regression
• Stacked Auto-Encoders (SAE)• Auto-Associative Neural Network
Dimensionality Reduction
• Convolutional Neural Network (CNN)• Deep Belief Networks (DBN)
• Deep Boltzmann Machine (DBM)
Clustering
Density Estimation
Applications
Image classification, Character recognition, Facial recognition, Surveillance systems
• The “Hey Siri” uses a Deep Neural Network (DNN) to convert the acoustic pattern of your voice into a probability distribution over speech sounds.
• It uses a temporal integration process to compute a confidence score (alpha values) that you uttered “Hey Siri”.
• If the score is high enough, Siri wakes up.
Example: Apple (The Detector: Listening for “Hey Siri”) 56
https://machinelearning.apple.com/2017/10/01/hey-siri.html
• Two major types of deep learning models
• Convolutional Neural Network (CNN)
• Recurrent Neural Network (RNN)
Deep Learning 57
• One to many: image captioning (ex, House, Dog, Trees)
• Many to one : Weather/Stock price forecasting• Many to many: video processing by frames to
caption, Language translation
• Usually applied for image recognition
• Regression : The output variable takes continuous values
• Classification : The output variable takes class labels
• Underneath it may still produce continuous values such as probability of belonging to a particular class.
Convolutional Neural Network (CNN) 58
http://cs231n.github.io/convolutional-networks/
Recurrent Neural Network (RNN)
• Similar to shallow RNNs, the deep RNNs are designed for modeling sequences and are capable of remembering past information and processing new events accordingly which was not possible with CNNs.
• Incredible success applying RNNs to a variety of problems:
• speech recognition
• language modeling
• translation
• image captioning
Recurrent Neural Network 60
<EOS> = end-of-sentence tag<BOS> = begin-of-sentence<pad> = Zeros are used, when there is no input at the time step
• Reinforcement learning is a general-purpose frame work for decision-making:
• An agent operates in an environment: Atari Breakout
• An agent has the capacity to act. Each action influences the agent’s future state
• Success is measured by a reward signal. Goal is to select action to maximize future reward
Deep Reinforcement Learning 61
https://deepmind.com/research/dqn/
Reward Discount factor State ActionNext Q-value
Mean Square
Popular Neural Network Architectures 62
Convolution Pooling Convolution Pooling Fully Connected Output Predictions
Dog(0.01)cat(0.04)
boat(0.94)bird(0.02)
Convolutional Neural Network
Recurrent Neural Network
LSTM LSTM
𝑦𝑡−1 𝑦𝑡 𝑦𝑡+1
• Deep Learning in Wireless Sensor Networks
• Network Traffic Classification
• Network Flow Prediction
• Deep Learning in Social Networks
• Mobility Prediction with Deep Learning
• Deep Learning in Cognitive Radio and Self-Organized Networks
• Deep Learning Based Routing
• Deep Learning in Internet-of-Things
• Deep Learning Approaches to Mobile Edge Computing
• Deep Learning for Network Security
Applications of Deep Learning in Network Related Areas 63
• TensorFlow is an open source software library for numerical computation using data flow graphs
• TensorFlow supports popular programming languages such as Python, C++, Java
• TensorFlow was originally developed by researchers and engineers working on the Google Brain Team within Google's Machine Intelligence research organization for the purposes of conducting machine learning and deep neural networks research, but the system is general enough to be applicable in a wide variety of other domains as well
Deep Learning implementation using open source: Use-Case 64
• Installing TensorFlow with native pip
Installing TensorFlow 65
If one of the following versions of Python is not installed on your machine, install using:Python 3.5.x 64-bit from python.orgPython 3.6.x 64-bit from python.org
C:\> pip3 install --upgrade tensorflow
C:\> pip3 install --upgrade tensorflow-gpu
https://www.tensorflow.org
Pip is a package management system used to install and manage software packages, such as those found in the Python Package Index.
TensorFlow visualization Tool: TensorBoard 66
• TensorFlow for training a massive deep neural network can be complex and confusing
• TensorBoard (visualization tools)• To make it easier to
• Understand
• Debug
• Optimize TensorFlow programs
• To visualize
• TensorFlow graph
• Plot quantitative metrics about the execution of graph
• Show additional data like images that pass through it
https://www.tensorflow.org
https://www.tensorflow.org/get_started/graph_viz
Supervised learning in Tensorflow 67
Training
Model
Output
Evaluation
Features
Extraction
Weight update
New
DATA
Labels
Raw
DATA
ModelFeatures
ExtractionOutput
Predicted
Outputs
Save Train Model
CNN,RNN, etc..
Training phase
Testing phase
• Build a convolutional neural network model to recognize the handwritten digits in the MNIST (Modified National Institute of Standards and Technology database) data set
1. Input Layers : monochrome 28x28 pixel images
2. Convolutional Layer 1: Applies 32 5x5 filters (extracting 5x5-pixel subregions), with ReLU activation function
3. Pooling Layer 1: Performs max pooling with a 2x2 filter and stride of 2 (which specifies that pooled regions do not overlap)
4. Convolutional Layer 2: Applies 64 5x5 filters, with ReLU activation function
5. Pooling Layer 2: Again, performs max pooling with a 2x2 filter and stride of 2
6. Dense Layer 1: 1,024 neurons, with dropout regularization rate of 0.4 (probability of 0.4 that any given element will be dropped during training)
7. Dense Layer 2 (Logits Layer): 10 neurons, one for each digit target class (0–9)
Example use-case: recognizing the handwritten digits 68
http://yann.lecun.com/exdb/mnist/ https://www.tensorflow.org/tutorials/layers
Convolutional Neural Network model for the use-case 69
InputOutput
• Input Layer
• batch_size : Size of the subset of example to use when performing gradient descent during training
• image_width : Width of the example images
• image_height : Height of the example images
• Channels : Number of color channels in the example images. For color images, the number of channels is 3 (red, green, blue). For monochrome images, there is just 1 channel (black)
• Here, MNIST dataset is composed of monochrome 28x28 pixel images, so the desired shape for our input layer is
• To convert our input feature map (features) to this shape, we can perform the following reshape operation:
• -1is used to reshape the input matrix into sequences
Input Layer: Convolutional Neural Network 70
TensorFLow function for Input layer:
input_layer = tf.reshape(features["x"], [-1, 28, 28, 1])
Feature
[batch_size, 28, 28, 1]
28 X28 784
• Convolutional layers, apply a specified number of convolution filters to the image
• In our use-case, we apply 32 5x5 filters to the input layer, with a ReLU activation function
• Kernel size is same as the filter size (5x5)
Training Model: Convolutional Layer 1 71
TensorFLow function for convolution:
conv1 = tf.layers.conv2d(
inputs=input_layer,
filters=32,
kernel_size=[5, 5],
padding="same",
activation=tf.nn.relu)
5X5X32 28 X28X32
Cov1 Output
output tensor produced by conv2d() has a shape of [batch_size, 28, 28, 32]
• Pooling layers downsample the image data extracted by the convolutional layers to reduce the dimensionality of the feature
• We apply the max_pooling2d() method in layers to construct a layer that performs max pooling with a 2x2 filter and stride of 2:
Training Model: Pooling Layer 1 72
TensorFLow function for Pooling:
pool1 = tf.layers.max_pooling2d(inputs=conv1, pool_size=[2, 2],
strides=2)
Output tensor produced by max_pooling2d() (pool1) has a shape of [batch_size, 14, 14, 32]
Pool1
14X14X32
Outputs ofPool1
2X2X32
1 2 5 6
3 4 7 8
strides=2
Training Model: Convolutional Layer 2 and Pooling Layer 2 73
Output pool2 has shape [batch_size, 7, 7, 64]
Cov2
14X14X64
Outputs of Cov2
5X5X64 2X2X64 7X7X64
Pool2Outputs of Pool2
• We connect a second convolutional and pooling layer to CNN using
• conv2d()
• max_pooling2d()
• For convolutional layer 2
• 64 5x5 filters with ReLU activation,
• For pooling layer 2,
• Same as pooling layer 1
• 2x2 max pooling filter with stride of 2
TensorFLow function for Convolution and Pooling:
conv2 = tf.layers.conv2d(
inputs=pool1,
filters=64,
kernel_size=[5, 5],
padding="same",
activation=tf.nn.relu)
pool2 = tf.layers.max_pooling2d(inputs=conv2, pool_size=[2, 2],
strides=2)
• Add a dense layer (1,024 neurons and ReLU activation)
• to perform classification on the features extracted by the convolution/pooling layers
• How to find number of neurons ? : Trials and errors
Training Model: Dense Layer 74
TensorFLow function for connecting Dense layer:
dense = tf.layers.dense(inputs=pool2_flat, units=1024, activation=tf.nn.relu)
Output tensor dropout has shape [batch_size, 1024]dense1024
ReLU activation
• The final layer is the logits layer, which will return the raw values for predictions
• Create a dense layer with 10 neurons (one for each target class 0–9), with linear activation
Training Model: Logits Layer 75
TensorFLow function for logits layer:
logits = tf.layers.dense(inputs= dense, units=10)
Final output tensor of the CNN, logits, has shape [batch_size, 10].Logits 10
0
9
• The logits layer of model returns predictions as raw values in a [batch_size, 10]-dimensional tensor
• The predicted class for each example: a digit from 0–9
Output: Generate Predictions 76
TensorFLow function for generating output
tf.argmax(input=logits, axis=1)
Training Loss Evaluation 77
Iterations
Trai
nin
g Lo
ss (
%)
Open source Machine Learning Library 78
https://www.svds.com/getting-started-deep-learning/
Applications of AI in 5G• Network Slicing Enablers• Wireless Network Virtualization• Evolution of Cellular Networks• Network slicing Deliverables • Network Slicing Industrial Efforts • AI and Network Slicing for 5G Networks: Use-Cases
79
5G cellular networks were assumed to be the key enabler and infrastructure provider in the ICT industry, by offering three types of services:
• Enhanced mobile broadband (eMBB)
• Ultra-reliable low latency service (URLLC)
• Massive machine-type communications (mMTC)
5G Promises
http://www.itu.int/dms_pubrec/itu-r/rec/m/R-REC-M.2083-0-201509-I!!PDF-E.pdf
80
• The existing mobile network architecture was designed to meet requirements for voice and conventional mobile broadband (MBB) services
• To meet the requirements of novel bandwidth hungry services, there is a need to deploy smarter 5G networks
• Note that, all novel services have very diverse requirements, thus having traditional RAN and core solutions for every service cannot guarantee the QoS
Challenges to realize 5G Networks 81
How to fulfil the diverse 5G networks requirements?
• Network slicing can fulfil the diverse requirements of these novel network services
• Network slicing enables one physical network into multiple, virtual, end-to-end (E2E) networks, each logically isolated including device, access, transport and core network
• A slice is dedicated for different types of service with different characteristics and requirements given to a Service End-point Agent (SEA)
• Enforce strong isolation between slices, i.e., actions in one slice do not affect another
Network Slicing
Physical Resource Pool
WindowsWindows
Virtualization
Slice 1WindowsWindowsSlice X
WindowsWindowsSlice Z
Network Slicing
82
• Network Slicing enablers: How to do it ?• Software-defined networking (SDN)
• Network Functions Virtualization (NFV)
• A single physical network will be sliced into multiple virtual networks:
• Different service types running across each virtual network, i.e., URLLC, EMBB, etc.
• Support different radio access networks (RANs), i.e., LTE, Wi-Fi, etc.
• It is envisaged that network slicing will be used to partition the core network and radio access networks
Network Slicing and 5G Networks 83
Network slicing enablers: Software defined network (SDN) 84
API to the data plane
(e.g., OpenFlow , ONOS)
Decentralized control plane
(which is closely tied to data
planes)
API to SDN application programmer
(who can now program the network as a system and not as a collection of
individual boxes)
Logically-centralized DP-
decoupled control
SDN Controller
At the highest level, the SDN movement is an effort to build networks you can
program at a higher level of abstraction— just as you can program a computer.
SDN enables programmability 85
Vertically integrated
Closed, proprietary
Slow innovation
Horizontal
Open interfaces
Rapid innovation
Specialized
Control
Plane
Specialized
Hardware
Specialized
Features
AppAppAppAppAppAppAppAppAppAppApp
Control
Plane
Control
Plane
Control
Plane or or
Open Interface
Merchant
Switching Chips
Open Interface
SDN enables programmability which is important for network slicing
• A network architecture concept that uses the technologies of IT virtualization to virtualize entire classes of network node functions that may connect, or chain together, to create communication services
• NFV is envisioned to play a crucial role in network slicing as it will be responsible to build isolated slices based on user service requirements
Network slicing enablers: Network function virtualization (NFV) 86
WindowsWindows
Physical Hardware
Virtualization
Virtual
Compute
WindowsWindowsVirtual
Storage
WindowsWindowsVirtual
Network
Mobility
Management
Entity (MME)
Serving Gateway
(S-GW)
Policy and
Charging Rules
Function (PCRF)
Man
agem
ent an
d O
rchestratio
n
• Due to massive success of NFV and SDN in wired domain, a number of studies are being conducted to adopt them both in the core and radio access networks (RANs) for future cellular networks such as:
• CORD (Central Office Re-architected as a Datacenter) [1]
• Radisys M-CORD [2]
• Wireless network virtualization (WNV) is a novel concept for virtualizing the RANs of future cellular networks
• WNV has a very broad scope ranging from spectrum sharing, infrastructure virtualization, to air interface virtualization
Wireless Network Virtualization 87
[1] https://opencord.org/[2] http://www.radisys.com/radisys-m-cord-open-platform-emerging-5g-applications
• WNV abstracts the physical wireless infrastructure and radio resources
• These resources are then isolated to a number of virtual resources (slices)
• The goal is to assign slices to different mobile virtual network operators (MVNOs) such that the network utility is maximized.
WNV: Slice Allocation 88
S.M Ahsan Kazmi, Choong Seon Hong, "A matching game approach for resource allocation in wireless network virtualization", The International Conference on Ubiquitous Information Management and Communication (IMCOM 2017), Jan. 05-07, 2017, Beppu, Japan
MVNO 1 MVNO 2 MVNO V
UE 1 UE 5UE 2UE 3 UE 4
UE k
Infrastructure Provider (InP)
2
FrequencyRad
io E
lem
ent
1
FrequencyRad
io E
lem
ent
C
FrequencyRad
io E
lem
ent
Mobile virtual network operators (MVNOs)
UE NUE n
UE l
User Equipment (UEs)
Slice (InP 1)
MVNO 1 MVNO 2 MVNO 3
UE 1 UE 4UE 2 UE 3 UE 5 UE 6
1 𝐶12 3
321 𝑆1
𝐶22 31
321
Slice (InP 2)
𝑆2
Physical Resource of InP 1
UE 7
Physical Resource of InP 2
• A practical deployment of a WNVinvolves a multi-cell scenario
• the coverage area will be servicedby a set of InPs
• The goal is efficient allocation of theslices such that the total performanceof WNV is improved.
• To solve this problem, we propose:
• Hierarchical matching game whichenables distributed implementationwhile satisfying efficient resourceallocation and strict isolation.
WNV: Service Selection and Resource Purchasing 89
S. M. Ahsan Kazmi, Nguyen H. Tran, Tai Manh Ho, Choong Seon Hong, "Hierarchical Matching Game for Service Selection and Resource Purchasing in Wireless Network Virtualization," IEEE Communications Letters (online)
Evolution of Cellular Networks
http://www.5gsummit.org/berlin/docs/slides/HansEinsiedler.pdf
90
• Ultra high bandwidth for enhanced mobile broadband (eMBB) through customized slice both at RAN and core.
• The core addresses this by placing the contents near to UE, i.e., mobile edge computing or smart caching schemes
• Ultra low delay/reliability for URLLC through customized slice both at RAN and CN
• Dedicated bandwidth allocation at core routers
Prospects of Network Slicing 91
• 3GPP working group on architecture (SA2) has already defined the basis for building an evolved core network
• The 5G infrastructure is expected to manage multiple slices on the same network infrastructure
• The envisioned architecture clearly differentiates between control plane (C-Plane) and user plane (U-Plane)
Role of 3GPP in Network Slicing
Third Generation Partnership Project (3GPP), “System Architecture for the 5G System,” 3GPP TS 23.501 v0.3.1, Mar. 2017.
92
PCFSMFs
eNB
gNB
Slice Manager
ng-RAN
NG2 Interface
UPF
AMF
NG15NG4
NG7 NG11
WindowsWindowsNetwork Slices
• In the control plane, new components are introduced to • Manage user authentication and registration (AMF)
• Support multiple connection sessions (SMF)
• Instruct different routing policies (PCF)
• The user plane is unified into a generic function (UPF) managing distinct data networks (DNs) through the next-generation-Radio Access Network (ngRAN)
• This new architecture allows for an easier network functions virtualization and enables flexible multitenant deployments
• RAN nodes (and functions) are virtualized and flexibly chained to provision end-to-end RAN slices with a dedicated SMF
Role of 3GPP in Network Slicing 93
Slicing: Next Generation Mobile Networks (NGMN)
https://www.ngmn.org/uploads/media/NGMN_5G_White_Paper_V1_0.pdf
94
CP: Control Plane
UP: User Plane
Vertical AP: Vertical Application
RAT: Radio Access Technology
Slicing: Huawei Technologies
http://www.huawei.com/minisite/5g/img/5G_Nework_Architecture_A_High_Level_View_en.pdf
95
RAN-RT: Radio Access Network-Real Time
RAN-NRT: Radio Access Network-non Real Time
AC: Access Cloud
CP: Control Plane
UP: User Plane
MCE: Mobile Cloud Engine
DC: Data Center
• Network Slicing certainly is one of the most discussed technologies these days. Network operators like KT, SK Telecom, China Mobile, DT, KDDI and NTT, and also vendors like Ericsson, Nokia and Huawei are all recognizing it as an ideal network architecture for the coming 5G era
• Ericsson has been working on network slicing with NTT DOCOMO since 2014. In 2016 the two announced a successful proof of concept of dynamic network slicing technology for 5G core networks
• They created a slice management function and network slices based on requirements such as latency, security or capacity
Network Slicing: Industrial Efforts 96
• In 2015 Ericsson and South Korea’s SK Telecom joined hands to develop and deploy network slicing technology optimized for 5G services.
• The two companies demonstrated the creation of virtual network slices optimized for services including super multi-view and augmented reality/virtual reality, massive Internet of Things and enterprise solutions.
• In November 2016, Huawei and Deutsche Telekom demonstrated the world’s first 5G end-to-end autonomous network slicing.
• The demo showed how different network slices can be created automatically and in an optimized way on a shared RAN, core and transport network.
Network Slicing: Industrial Efforts 97
What is missing?
• The basic goal of an AI in 5G is its ability to extract, predict, and characterize specific patterns from datasets
• To unleash the true potential of 5G networks:• Intelligent functions using AI across both the edge and core of the network are
required along with the novel enabling technologies
• AI functions must be able to:• Adaptively exploit the wireless system resources • Generated data to optimize network operation• Guarantee the QoS in real time
• Such mobile edge and core intelligence can only be realized by integrating fundamental notions of artificial intelligence (AI) across the wireless infrastructure and end-user devices
AI for 5G Networks 98
• Role of AI in 5G networks: Exploit big data analytics to enhance situational awareness and overall network operation such as:
• Fault Monitoring
• User Tracking
• Cell Association
• Radio Resource Management
• Cache Resource Management
• Mobility Management
• Management and Orchestration
• Service Provisioning Management
• And so on..
AI for 5G Networks
AI-based system operation is no longer a privilege, but rather a necessity for 5G and beyond networks.
99
AI for Networking Slicing in 5G Networks
Perception
LearningProblem Solving
Reasoning
Core Network
Gateway Switch
Radio Access Network
RBsCache
Slice Manager: AI Enabled
mMTCeMBB URLLC
100
• AI engine can act as an application on top of slicing manager or run as an independent network entity, and communicate with RAN and CN
• AI enabled Slice manager related information:
• The slice manager will read service-level agreements of Users, e.g., requirements on rate, coverage, failure duration, redundancy, etc.
• UE-level information such as receiver category, battery limitation, power, mobility and etc.
• Network-level information such as resource spectrum, number of serving users, QoS (quality of service), key performance indicators of network functions, scheduled maintenance period, and etc.
• Infrastructure-level information such as server type, CPU, memory, storage, and etc.
AI enabled 5G networks 101
Information sent to the AI enabled Slice Manager
Perception
LearningProblem Solving
Reasoning
RBs CacheSlice Manager: AI Enabled
102
RAN Agent
Controller API
Radio Access Network
Core Network
Gateway Switch
Core Network Agent
Spectrum, Requirements on rate, coverage, etc.
Service-level agreements
Moving speed, power, etc.
UE level information
CPU, memory, network capabilities, etc.
Infrastructure level information
Number of serving users, etc.
Network level information
• These information can be acquired via SDN controllers or open API
• Then, the slice manager will utilize its embedded modules of problem solving to process the obtained information, and feedback learning results
• These results can include:• Traffic characteristic analysis reports such as service provisioning suggestion to CN
or RAN
• User-specific controlling information such as serving priority, bandwidth allocation, mobility tracking command to RAN
• Network configuration notification such as parameter adjustment, access method, network error alert to the controllers of RAN or CN
AI enabled 5G networks 103
Perception
LearningProblem Solving
Reasoning
Physical Resource Blocks (PRBs)Cache
Slice Manager: AI Enabled
104
RAN Agent
Radio Access Network
Core Network
Gateway Switch
Core Network Agent
PRBs adjustment, cache adjustment, etc.
Service Provisioning
serving priority, PRB allocation, etc.
UE specific information
Adjust CPU and Storage, assign network capabilities , etc.
Infrastructure level information
Adjustment in number of serving users, etc.
Network level service provisioning
Controller API
Information sent by the AI enabled Slice Manager
• Currently network slices are created and modified based on pre-defined stimuli, i.e., time of day, scheduled maintenance, social event, and etc.
• Through AI enabled slice manger, we can perform prediction of the UEs requirements based on service-level agreements of UEs, density of UEs, and QoS of UEs
• These prediction results can then be used in deciding new creation or deletion of network slices. Moreover, we can also determine the types of slices required in the network
• Thus, the slice manager can decide and inform the RAN and CN to create or delete slices of specific types
Use-Case: Slice Creation 105
• Goal : Maximize the cache hit, in order to reduce access latency for URLLC service
• Potential Benefits : Enhanced cache hit, low access latency, bandwidth saving for backhaul
• Approach: Deep Learning using the Recurrent Neural Network
• Input : Sequences of content access information
• Output : Content popularity values and cache decision
• Learning Model :
• Long Short Term Memory (LSTM)
• Gated Recurrent Unit (GRU)
• Simple RNN
Use-Case: Content’s Popularity prediction and Cache decision 106
Content’s Popularity prediction and Cache decision 107
Feed Collected Data
Learning Content’s
Popularity with Deep Recurrent Neural Network
Cache Decision
Performance Measurement
AI enabled Slice Manager
Cache
a b c d
Cache
e f g a
𝑥1 𝑥2 𝑥𝑛
RNN
Cache Decision
Performance Measurement
Up
dat
e W
eigh
t
Sequences Inputs : Features
Output: Popularity Score
Long Short Term Memory (LSTM)
Gated Recurrent Unit (GRU)
Simple Recurrent Neural Network
• Type of content• Rating (IMDB)• Time Zone• Hit count• User ID• movID• Quality of video
Sequence Inputs: Features
RNNRNN
𝑦
Training
Prediction
• Dataset : MovieLens is a popular dataset for video content recommendation, i.e., https://grouplens.org/datasets/movielens/. We divide the dataset into training (70%) and test data (30%) sets.
• We build the RNN model to predict the Content’s Popularity in TensorFlow. We have two phases.
• Training: • Batch size =200, i.e., movie context. This is the input of the RNN model.• Hidden layers = 200, we use the Long Short Term Memory (LSTM).• Time frame = 3, i.e., after every three sequences of input, the model will predict.• Loss = root mean squared error (RMSE), once the model predicts, we calculate the loss
using RMSE.• Optimizer = adam, we use the adam optimizer to minimize the loss.• Accuracy = We define the accuracy threshold for stopping the learning phase.
• Testing: • We evaluate the accuracy of the test data from the trained model.
Dataset Preprocessing and Model Description 108
Implementing our model using TensorFlow 109
RNN RNN RNN
RNN RNN RNN
RNN RNN RNN
𝑌1
𝑋𝑡1
𝑋𝑡2
𝑋𝑡3
Inputs Layer Hidden 1 Hidden N
𝑌1
𝑅𝑀𝑆𝐸
Movie ID, Number of requests, Region,
Type of movie.
Context Info
Pop
ularity Sco
res
True Label
Output: Popularity Score
Trained Model
𝑋𝑡
𝑋𝑡+1
Trai
n D
ata
Test
Dat
a
Dat
a se
t
𝑌1 𝑌1
𝑅𝑀𝑆𝐸 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦
Cache Decision
Hidden layers = 200Batch size =200 (number of Contents)Time frame = 3Activation = sigmoidOptimizer = adamLoss = mean square error
𝑋𝑡+2
Prediction Results 110
• Mobility Management
• Fault Monitoring
• User Tracking
• Cell Association
• Load Balancing
• Power Management
• Radio Resource Management
• Cache Resource Management
• And many more benefits..
What more can be achieved via enabling AI? 111
• The use of artificial intelligence will play a vital role for enabling a variety of applications in 5G and beyond wireless networks.
• AI definitively provides precious opportunities to analyze trends and recognize patterns. However, it is difficult to perfectly predict the desired results by using traditional simple models such as shallow ANNs.
• Deep Neural Networks are envisioned to fill this gap and serve as key predicting enabler to support the 5G networks
• Network slicing coupled with AI will be defining the future of wireless networks.
Conclusion 112