1 of 62

Choong Seon Hong

School of Computing

Kyung Hee University, Republic of Korea

URL: http://networking.khu.ac.kr

AI for Networking

NETWORKING

INTELLIGENCE LAB

2 of 62

Outline

  • Introduction
    • Evolution of AI
    • Myths on AI
    • Why AI is Important in Networking?
  • AI Research Activities for Networking
    • AI Growth in Telecommunication Industry
    • Ongoing Research in Academic
  • Activities in Networking Intelligence
    • DRL for Aerial and IRS Networking
    • XAI for Internet of Everything
    • Federated, Distributed and Democratized Learning for Networking Environments
  • Future Direction on AI for Networking
    • Explainable Multi-Modal AI for Networking
    • Distributed Edge AI
    • Democratized Learning
    • Digital Twin for AI
    • Semantic AI Networking
    • Potential AI Research Topics in IRTF

2/62

NETWORKING

INTELLIGENCE LAB

3 of 62

Introduction: The Evolution of Artificial Intelligence (AI)

Modified from Source: https://twitter.com/mikequindazzi/status/835589969909424130

1950

Alan Turing Proposes the Turing Test

1950

Isaac Asimov proposes the Three Laws of Robotics

1951

First AI based Program was written

1955

First self learning game playing program is written

1959

MIT AI Lab

is setup

1961

First Robot is inducted into GM’s assembly

production line

1963

First Machine Learning program is written

1964

First demonstration of an AI program which understand Natural Language

1965

First AI based Chat-bot (ELLZA) was created

1969

Stanford Research Institute (SRI) demonstrates the first locomotive and intelligent robot (Shakey)

1969

First autonomous vehicle is created at the Stanford AI LAB

1974

First rule-based AI expert system for medical diagnostics

1980

LISP based machines are developed and marketed

1986

Learning representations by back-propagating error

(Backpropagation)

[G. Hinton]

1997

IBMs Deep Blue beats Gary Kasparov at Chess

1998

Convolutional Neural Network is introduced by Yann LeCun

2004

DAPRA introduces the first challenge for Autonomous Vehicles

2005

AI based recommendation engines

2009

Google builds Self Driving Car

2010

Narrative Science’s AI demonstrates ability to write reports

2011

IBM Watson beats Jeopardy champions

2011

Personal Assistants like Siri, Google Now and Cortana become mainstream

2015

Elon Musk and others announce a $1B nonprofit open source initiative, OPEN AI to develop friendly AI

2016

Google’s DeepMind AlphaGo defeats Go’s champions

2016

NVIDIA announces supercomputer for Deep Learning and AI

2017

AlphaGo Zero which learns from scratch;

Federated Learning�collaboratively trains a global model

2018

OpenAI 5

(Dota 2)

2019

OpenAI 5

(Defeat the world’s top Dota 2 team)

1999

First Emotional AI machines demonstrated at MIT AI Lab

1990

Probabilistic models of sequences

(Yoshua Bengio)

Turing Award 2018

2020

Open AI’s GPT-3

DeepMind’s AlphaFold V2.0

2021

Multimodal

Neurons in ANN by Open AI

1958

Introducing the neural networks perceptron model

[Frank Rosenblatt]

1948

Warren McCulloch &

Walter Pitts: Modelling the first neural network using electric circuit

Neural Networks

2022

Open AI’s GPT- 4 (?),

Low-code and no-code AI

3/62

NETWORKING

INTELLIGENCE LAB

4 of 62

Introduction: AI Applications and Myths on AI

Source: https://www.javatpoint.com/application-of-ai

Source: https://ai.google/static/documents/exploring-6-myths.pdf

  • AI, machine learning, and deep learning are all the same thing
  • All AI systems are “black boxes,” far less explainable than non-AI
  • AI systems are only as good as the data they train on
  • AI systems are inherently unfair
  • AI will make human labor obsolete
  • AI is approaching human intelligence

Typical AI application domain

Myths on AI

4/62

NETWORKING

INTELLIGENCE LAB

5 of 62

Introduction: Why AI is Important in Networking?

Emerging Use Cases of AI in Networking

​

​

5G, 6G, Visible Light Communication, etc

Emerging Communication Technologies

5/62

NETWORKING

INTELLIGENCE LAB

6 of 62

AI Research Activities for Networking

6/62

NETWORKING

INTELLIGENCE LAB

7 of 62

AI Research Activities for Networking: AI Growth in Telecommunication Industry

  • Artificial intelligence (AI) represents one of the biggest emerging opportunities in technology
    • By the end of 2027, the global AI in telecommunication market is expected to reach an impressive $14.99B to grow a CAGR of 42.6% [1]

$14.99B

$1.189B

[1] Source: https://www.researchandmarkets.com/reports/4375395

[2] Source: https://www.n-ix.com/ai-in-telecommunications/

Global AI Market Size [1]

Global AI market size in telecommunication [2]

CAGR: compound annual growth rate

7/62

NETWORKING

INTELLIGENCE LAB

8 of 62

AI Research Activities for Networking: Ongoing Research in Academic

[1] Modified from Source: Y. Siriwardhana, P. Porambage, M. Liyanage and M. Ylianttila, "AI and 6G Security: Opportunities and Challenges," 2021 Joint European Conference on Networks and Communications & 6G Summit (EuCNC/6G Summit), June 2021.

[2] https://petarpopovski-51271.medium.com/seven-briefs-on-semantic-communication-and-6g-693c35600148

AI and 6G

The augmented triangle of 6G [2]

Status

Detection

8/62

NETWORKING

INTELLIGENCE LAB

9 of 62

AI Research Activities for Networking: Ongoing Research in Academics

[1] M. K. SHEHZAD, L. Rose, M. M. Butt, I. Z. Kovacs, M. Assaad and M. Guizani, "Artificial Intelligence for 6G Networks: Technology Advancement and Standardization," in IEEE Vehicular Technology Magazine (Early Access), May 2022.

[2] L. U. Khan, W. Saad, Z. Han, E. Hossain and C. S. Hong, "Federated Learning for Internet of Things: Recent Advances, Taxonomy, and Open Challenges," in IEEE Communications Surveys & Tutorials, vol. 23, no. 3, pp. 1759-1799, third quarter 2021

AI/ML Challenges for Networking Research

  • K-means
  • PCA
  • Mixture Models
  • Auto Encoder
  • GAN

9/62

NETWORKING

INTELLIGENCE LAB

10 of 62

AI Research Activities for Networking: Ongoing Research in Academic

  • CNN and DNN
    • Intrusion detection
    • Network traffic prediction
    • Power control
    • Routing wireless backbone

Role of Neural Networks-based AI Models for Solving Networking Challenges

  • Auto encoder and GAN
    • Intrusion detection and prevention
    • Spectrum sensing and reconstruction
    • Channel estimation and error correction
    • Localization
    • Beamforming and tracking by reconstruction

Convolutional neural network (CNN)

Deep neural networks (DNN)

Generative adversarial network (GNN)

  • It depends on the pre-trained AI model
  • It can classify based on the label
  • It can not capture dynamic changes that are not in training data
  • It can not handle time depended decisions
  • RNN, GRU, and LSTM
    • Network resource slicing
    • Indoor and outdoor positioning
    • Channel estimation
    • Time series intrusion detection

​

  • It can solve the time dependent problem
  • It can not capture dynamic changes that are not in training data
  • It cannot reconstruct
  • It can reduce the data dimension
  • Can ignore noise in data during training and execution
  • It can reconstruct and suitable for auto correction
  • It can not capture dynamic changes

Long short-term memory (CNN)

Gated recurrent unit (GRU)

Recurrent neural network (RNN)

10/62

NETWORKING

INTELLIGENCE LAB

11 of 62

AI Research Activities for Networking: Ongoing Research in Academic

Experience

Policy/Value

Action

Networking

Environment

Model free RL

Experience

Policy/Value

Action

Networking

Environment

Updating Model(trained)

Model-based RL

Learning outputs

  • A model-free algorithm either estimates a value function or the policy directly from experience
  • After learning, the agent can make predictions about what the next state and reward will get before it takes each action
  • M. S. Munir, S. F. Abedin, N. H. Tran and C. S. Hong, "When Edge Computing Meets Microgrid: A Deep Reinforcement Learning Approach," in IEEE Internet of Things Journal, vol. 6, no. 5, pp. 7360-7374, Oct. 2019.
  • M. S. Munir, D. H. Kim, S. W. Kang, L. Zou and C. S. Hong, "Intelligent Grid Shepherd: Towards a Resilient Distributed Energy Resources Control System," 2021 22nd Asia-Pacific Network Operations and Management Symposium (APNOMS), 2021.

Role of Reinforcement Learning for Solving Networking Challenges

Model free RL

11/62

NETWORKING

INTELLIGENCE LAB

12 of 62

AI Research Activities for Networking: Ongoing Research in Academic

Role of Reinforcement Learning for Solving Networking Challenges

Reinforcement learning (RL)

Model free RL

Model-based RL

On-Policy (policy optimization)

Off-Policy (value/Q- function)

Model learn (updating)

Model given (fixed)

  • Content caching
  • UAV deployment and control
  • Power control
  • Service placement
  • Handover
  • Routing
  • Networking and computational resource management
  • Network resource slicing
  • Indoor and outdoor positioning
  • Channel estimation and error correction
  • Intrusion detection and prevention
  • Localization
  • Spectrum sensing and reconstruction
  • Beamforming and tracking by reconstruction
  • A2C/A3C
  • PPO
  • Q-learning
  • DQN
  • AlphaZero
  • World model

12/62

NETWORKING

INTELLIGENCE LAB

13 of 62

  • Activities in Networking Intelligence
  • DRL for Aerial and IRS Networking

13/62

NETWORKING

INTELLIGENCE LAB

14 of 62

DRL for Aerial Networking: Data Freshness and Energy-Efficient UAV Navigation

System model for heterogeneous unmanned aerial networks with edge computing

S. F. Abedin, M. S. Munir, N. H. Tran, Z. Han and C. S. Hong, "Data Freshness and Energy-Efficient UAV Navigation Optimization: A Deep Reinforcement Learning Approach," in IEEE Transactions on Intelligent Transportation Systems, vol. 22, no. 9, pp. 5994-6006, Sept. 2021.

  • For enabling computation-oriented communications (COC) applications such as virtual and augmented reality (VR and AR), real-time monitoring, and surveillance.

14/62

NETWORKING

INTELLIGENCE LAB

15 of 62

DRL for Aerial Networking: Data Freshness and Energy-Efficient UAV Navigation

  • Motivations
    • How to maximize energy efficiency by optimizing the UAV-BS trajectory policy?
    • How to incorporate different contextual information such as energy and age of information (AoI) constraints to ensure the data freshness at the ground BS?

​

  • Designing a navigation policy for multiple unmanned aerial vehicles (UAVs)
    • Where mobile base stations (BSs) are deployed to improve the data freshness and connectivity to the Internet of Things (IoT) devices
    • We incorporate different contextual information such as energy and age of information (AoI) constraints to ensure the data's freshness at the ground BS

​

  • Then, we propose an agile deep reinforcement learning with an experience replay model to solve the formulated problem
    • That is concerning the contextual constraints for the UAV-BS navigation

​

​

​

S. F. Abedin, M. S. Munir, N. H. Tran, Z. Han and C. S. Hong, "Data Freshness and Energy-Efficient UAV Navigation Optimization: A Deep Reinforcement Learning Approach," in IEEE Transactions on Intelligent Transportation Systems, vol. 22, no. 9, pp. 5994-6006, Sept. 2021.

Summary of investigations

15/62

NETWORKING

INTELLIGENCE LAB

16 of 62

DRL for Aerial Networking: Data Freshness and Energy-Efficient UAV Navigation

S. F. Abedin, M. S. Munir, N. H. Tran, Z. Han and C. S. Hong, "Data Freshness and Energy-Efficient UAV Navigation Optimization: A Deep Reinforcement Learning Approach," in IEEE Transactions on Intelligent Transportation Systems, vol. 22, no. 9, pp. 5994-6006, Sept. 2021.

  • UAV-BS navigation optimization problem under contextual constraints (i.e., trajectory, AoI, energy efficiency constraints):
  • Decision sub-set of trajectory

Non-overlapping trajectories of the UAV-BSs except the ground BS where information update occurs

  • Total energy efficiency for UAV-BS:

Back-haul channel capacity between UAV-BS and ground BS

Channel capacity between IoT and UAV-BS

Transmission energy of UAV-BS to ground BS

UAV-BS mobility energy cost

A set of given trajectory points

Joint trajectory configuration of the UAV-BSs for all the trajectory points are covered interdependently

Assuring the total energy efficiency of the UAV-BSs

Ensuring an average freshness of information updates by configuration

16/62

NETWORKING

INTELLIGENCE LAB

17 of 62

DRL for Aerial Networking: Data Freshness and Energy-Efficient UAV Navigation

Proposed Trajectory Policy Algorithm Based On Deep Q-learning

S. F. Abedin, M. S. Munir, N. H. Tran, Z. Han and C. S. Hong, "Data Freshness and Energy-Efficient UAV Navigation Optimization: A Deep Reinforcement Learning Approach," in IEEE Transactions on Intelligent Transportation Systems, vol. 22, no. 9, pp. 5994-6006, Sept. 2021.

Learning and Store the Q-network

Execution and update Q-network

17/62

NETWORKING

INTELLIGENCE LAB

18 of 62

DRL for Aerial Networking: Data Freshness and Energy-Efficient UAV Navigation

Performance Evaluation

S. F. Abedin, M. S. Munir, N. H. Tran, Z. Han and C. S. Hong, "Data Freshness and Energy-Efficient UAV Navigation Optimization: A Deep Reinforcement Learning Approach," in IEEE Transactions on Intelligent Transportation Systems, vol. 22, no. 9, pp. 5994-6006, Sept. 2021.

Avg. AoI vs. no. of trajectory

Avg. EE vs. no. of trajectory

Avg. bandwidth vs. no. of Trajectory

Avg. EE vs. AoI threshold

Simulation settings

18/62

NETWORKING

INTELLIGENCE LAB

19 of 62

DRL for Efficient Networking: RIS Controlling for 6G Network

System model for multiple RISs-enabled downlink communication networks

P. S. Aung, Y. K. Tun, Z. Han. and C. S. Hong, “Energy-Efficiency Maximization of Multiple RISs-Enabled Communication Networks by Deep Reinforcement Learning," IEEE International Conference on Communications 2022 (ICC 2022), May 2022.

.

  • Motivation
    • Recently, in 5G and upcoming 6G cellular networks, reconfigurable intelligent surface (RIS) has perceived a prodigious interest in both academic and industrial fields due to
      • Easy deployment
      • Spectral efficiency enhancement
      • Cost-effectiveness

19/62

NETWORKING

INTELLIGENCE LAB

20 of 62

DRL for Efficient Networking: RIS Controlling for 6G Network

  • We first investigate the energy-efficient multiple RISs-enabled downlink communication system from the base station (BS) to users

​

  • We then formulate the joint optimization problem of user-RIS association, reflective elements ON/OFF states, phase shift, and transmit power in order to maximize energy-efficiency

​

  • To address this challenge of a mixed-integer and non-convex problem,
    • We decompose the formulated problem into sub-problems
      • Joint user-RIS association, reflective elements ON/OFF states, and phase shift problem
      • Power control problem

​

  • A DRL approach and convex optimization technique are proposed to solve the sub-problems, alternatively

Summary of Investigations

P. S. Aung, Y. K. Tun, Z. Han. and C. S. Hong, “Energy-Efficiency Maximization of Multiple RISs-Enabled Communication Networks by Deep Reinforcement Learning," IEEE International Conference on Communications 2022 (ICC 2022), May 2022.

.

20/62

NETWORKING

INTELLIGENCE LAB

21 of 62

DRL for Efficient Networking: RIS Controlling for 6G Network

The goal of this work is to maximize the energy efficiency of the system by jointly optimizing the user-RIS association, reflective elements ON/OFF states, RIS phase shift, and the transmit power of the BS.

  • The energy efficiency maximization problem for multiple RISs-enabled communication networks is as follows:

Achievable data rate

To guarantee that the total transmit power of the BS is less than the maximum available power

QoS constraint of each user

The phase shift values should be between 0 to 2π

 

Association

Reflective elements ON/OFF states

Power

P. S. Aung, Y. K. Tun, Z. Han. and C. S. Hong, “Energy-Efficiency Maximization of Multiple RISs-Enabled Communication Networks by Deep Reinforcement Learning," IEEE International Conference on Communications 2022 (ICC 2022), May 2022.

.

Phase shift

Power consumption

Binary constraints of the user-RIS association and reflective elements ON/OFF states

21/62

NETWORKING

INTELLIGENCE LAB

22 of 62

DRL for Efficient Networking: RIS Controlling for 6G Network

Proposed solution

  • Each state:
  • User-RIS association
  • Reflective elements ON/OFF states
  • Phase shift variable
  • Channel gain of the direct link between the BS and each user
  • Channel gain between the BS and RIS
  • Channel gain between the RIS and each user
  • Transmit power control

P. S. Aung, Y. K. Tun, Z. Han. and C. S. Hong, “Energy-Efficiency Maximization of Multiple RISs-Enabled Communication Networks by Deep Reinforcement Learning," IEEE International Conference on Communications 2022 (ICC 2022), May 2022.

.

DRL for joint user-RIS association, reflective elements ON/OFF states and phase shift

  • Each action:

22/62

NETWORKING

INTELLIGENCE LAB

23 of 62

DRL for Efficient Networking: RIS Controlling for 6G Network

Simulation Results

P. S. Aung, Y. K. Tun, Z. Han. and C. S. Hong, “Energy-Efficiency Maximization of Multiple RISs-Enabled Communication Networks by Deep Reinforcement Learning," IEEE International Conference on Communications 2022 (ICC 2022), May 2022.

.

Convergence of reward function over different time steps

Comparison of sum-rate for different transmit power

Comparison of energy efficiency for different

transmit power

CDF of sum-rate for different number of reflecting elements

23/62

NETWORKING

INTELLIGENCE LAB

24 of 62

  • Activities in Networking Intelligence
  • XAI for Internet of Everything

24/62

NETWORKING

INTELLIGENCE LAB

25 of 62

Explainable Artificial Intelligence for Internet of Everything

Background and Motivation

[1] AI EXPRESS, “A General Guide to Internet of Everything (IoE)”, https://aiexpress.io/a-general-guide-to-internet-of-everything-ioe, December 26, 2021.

​

Four pillars of Internet of Everything (IoE) [1]

Why IoE?

  • The challenges incorporate when the AI model conceives a lake of interpretation and intuition to the network service provider

​

  • For assuring the quality of IoE services delivery, we must have to analyze the contextual metrics of IoE
      • User speed or mobility
      • Decisions based on data
      • Download and upload processes
      • CQI, RSRP, RSRQ, and SINR

​

​

​

CQI: channel quality index

RSRP: reference signal received power

RSRQ: reference signal received quality

SINR: signal to interference and noise ratio

25/62

NETWORKING

INTELLIGENCE LAB

26 of 62

Explainable Artificial Intelligence for Internet of Everything

A system model of explainable artificial intelligence-enabled quality-aware IoE service delivery

  • The sixth-generation (6G) wireless networks will need sophisticated artificial intelligence (AI) to automate information delivery simultaneously for mass autonomy, human-machine interfacing, and targeted mission-critical services [1]

[1] W. Guo, “Explainable Artificial Intelligence for 6G: Improving Trust between Human and Machine,” IEEE Communications Magazine, vol. 58, no. 6, pp. 39-45, June 2020.

[2] Matt Turek, “Explainable Artificial Intelligence (XAI)”, https://www.darpa.mil/program/explainable-artificial-intelligence, Visited March 16, 2022.

[3] M. S. Munir, S. B. Park, and C. S. Hong, “An Explainable Artificial Intelligence Framework for Quality-Aware IoE Service Delivery," IEEE International Conference on Communications 2022 (ICC 2022), May 2022.

  • Why did you do that?
  • Why not something else?
  • When do you succeed?
  • When do you fail?
  • When can I trust you?
  • How can I correct an error?
  • I understand why
  • I understand why not
  • I know when you succeed
  • I know when you fail
  • I know when to trust you
  • I know why you erred

Modified source: https://www.darpa.mil/program/explainable-artificial-intelligence

26/62

NETWORKING

INTELLIGENCE LAB

27 of 62

Explainable Artificial Intelligence for Internet of Everything

  • Introducing a quality-aware IoE service delivery mechanism by proposing a new explainable artificial intelligence framework for EPC core

​

  • Formulating a quality-aware IoE service delivery problem for EPC core,
      • The objective is to maximize the channel quality index (CQI) of each IoE service user

​

  • A multivariant regression problem is devised to solve the formulated problem,
      • Where explainable coefficients of the contextual matrices are estimated by Shapley value interpretation

​

  • XAI-enabled quality-aware IoE service delivery algorithm is developed and implemented
      • That can ensure the interpretation of contextual relationships among the metrics to reconfigure network parameters

Summary of key contributions

M. S. Munir, S. B. Park, and C. S. Hong, “An Explainable Artificial Intelligence Framework for Quality-Aware IoE Service Delivery," IEEE International Conference on Communications 2022 (ICC 2022), May 2022.

Shapley value: The extent of contribution from contextual parameters

27/62

NETWORKING

INTELLIGENCE LAB

28 of 62

Explainable Artificial Intelligence for Internet of Everything

M. S. Munir, S. B. Park, and C. S. Hong, “An Explainable Artificial Intelligence Framework for Quality-Aware IoE Service Delivery," IEEE International Conference on Communications 2022 (ICC 2022), May 2022.

Problem Formulation of the Quality-Aware IoE Service Delivery

  • The quality of each IoE service completely relies on correlation among
  • The rewritten set of contextual matrices is represented as,

{Speed, RSRP, RSRQ, SINR, RSSI, CQI, Downlink, Uplink}

Ensuring minimum level of RSRP

Association variable

Downlink data rate

Uplink data rate

Objective is to maximize the channel quality index (CQI) of each IoE service user using contextual parameters

Assuring minimum level of RSRQ

Capturing service user mobility

Establishing correlation among the contextual matrices

Ensuring each IoE service user belongs to signal range

Why are changes required to enhance CQI of each IoE user ?

N: total number of features

K: number of user

B: number of BS or gNB

CQI

28/62

NETWORKING

INTELLIGENCE LAB

29 of 62

Explainable Artificial Intelligence for Internet of Everything

M. S. Munir, S. B. Park, and C. S. Hong, “An Explainable Artificial Intelligence Framework for Quality-Aware IoE Service Delivery," IEEE International Conference on Communications 2022 (ICC 2022), May 2022.

  • Minimizing regression loss while considering the Shapley value coefficients for contextual interpretation
  • The regression problem is based on the coefficients of contextual features:
  • Coefficients(Sharpley value) of contextual features:

denotes the intercept and

where

represents contextual input

The objective is to minimize the loss while considering the Shapley value coefficients for contextual interpretation

*intercept: threshold

29/62

NETWORKING

INTELLIGENCE LAB

30 of 62

Explainable Artificial Intelligence for Internet of Everything

M. S. Munir, S. B. Park, and C. S. Hong, “An Explainable Artificial Intelligence Framework for Quality-Aware IoE Service Delivery," IEEE International Conference on Communications 2022 (ICC 2022), May 2022.

  • User speed, RSRP, RSRQ, SINR, RSSI, CQI, downlink bit rate, and uplink bit rate form the historical data

​

  • Initializing parameters for no. of RB, coefficients for contribution of all contextual matrices, coalition with other features (i.e., players), intercept

​

  • Output: association, downlink and uplink data rates, contextual coefficients
  • Calculation and estimation of RSRP, RSRQ, SINR, CQI, and contextual coefficients (Shapley value) for all service users
  • Training of the proposed explainable AI-based multi-variant regression model for quality-aware IoE service delivery
  • Minimizing regression loss
  • Decision making, explanation, and error correction (reconfiguration) to meet the CQI
  • The complexity of Algorithm 1 leads to :

N: total number of features

K: number of users

B: number of BS or gNB

30/62

NETWORKING

INTELLIGENCE LAB

31 of 62

Explainable Artificial Intelligence for Internet of Everything

M. S. Munir, S. B. Park, and C. S. Hong, “An Explainable Artificial Intelligence Framework for Quality-Aware IoE Service Delivery," IEEE International Conference on Communications 2022 (ICC 2022), May 2022.

Experiment Setup

Fig. 4. Considered topology for evaluating the proposed XAI-enabled IoE service delivery framework based on dataset (B 2020:02:13 13:03:24) [11]

  • EPC core computational server
    • A Core i9 processor (2.8 GHz) along with 64 GB of RAM
    • Implemented on Python framework

31/62

NETWORKING

INTELLIGENCE LAB

32 of 62

Explainable Artificial Intelligence for Internet of Everything

Performance Evaluation

M. S. Munir, S. B. Park, and C. S. Hong, “An Explainable Artificial Intelligence Framework for Quality-Aware IoE Service Delivery," IEEE International Conference on Communications 2022 (ICC 2022), May 2022.

Achieved higher CQI due to fairly controlled Shapley value-based prominent features (coefficients) by the XAI framework

Improvement rate up to:

  • Downlink: 42.43%
  • Uplink: 28.57%

Trend analysis on distinct topologies

Due to mobility

Due to channel condition

Maximum correlation between SINR and RSRQ(0, -2) : 27%

32/62

NETWORKING

INTELLIGENCE LAB

33 of 62

  • Activities in Networking Intelligence
  • Federated, Distributed and Democratized Learning for Networking Environments

33/62

NETWORKING

INTELLIGENCE LAB

34 of 62

Edge-Based Federated Learning

  • The under-explored resource allocation for the Federated Learning scheme:
    • The uncertainty of wireless channels
    • UEs with heterogeneous power constraints
    • The difference in local training data size

​

  • Contributions of [1]:
    • Formulate a Federated Learning over wireless network problem, namely (FEDL)
    • Decompose the non-convex FEDL problem and transform it to three convex sub-problems and obtain the globally optimal solution
    • Computation and communication latencies determined by learning accuracy level
    • Trade-off between the Federated Learning time and UE energy consumption.

​

​

​

[1] Nguyen H. Tran, Wei Bao, Albert Zomaya , Minh N.H. Nguyen and Choong Seon Hong, “Federated Learning over Wireless Networks: Optimization Model Design and Analysis,” IEEE International Conference on Computer Communications (INFOCOM 2019)

[2] Dinh, Canh T., Nguyen H. Tran, Minh NH Nguyen, Choong Seon Hong, Wei Bao, Albert Y. Zomaya, and Vincent Gramoli. "Federated learning over wireless networks: Convergence analysis and resource allocation." IEEE/ACM Transactions on Networking 29, no. 1 (2020): 398-409.

Motivation and Contributions

34/62

NETWORKING

INTELLIGENCE LAB

35 of 62

Edge-Based Federated Learning

  •  

​

Step 3. Update Global Model: The local model parameters and gradients are aggregated at the controller

​

​

​

Step 4. Update Learning Parameters: These updated learning parameters then are broadcast to all UEs.

Until a global error is achieved.

Federated Learning Scheme

Iterative Process

1. Local computation

2. Transmit Learning Parameters

3. Update Global Model

UE 1

UE 2

UE 3

4. Update Learning Parameters

FL Mechanism

35/62

NETWORKING

INTELLIGENCE LAB

36 of 62

Edge-Based Federated Learning

​

    • Minimizing the all UEs’ energy consumption and learning time

​

Communication Time

Computational Time

CPU cycle of UEs

Transmission power

Local error

FEDL optimization problem

[1] Nguyen H. Tran, Wei Bao, Albert Zomaya , Minh N.H. Nguyen and Choong Seon Hong, “Federated Learning over Wireless Networks: Optimization Model Design and Analysis,” IEEE International Conference on Computer Communications (INFOCOM 2019)

36/62

NETWORKING

INTELLIGENCE LAB

37 of 62

Edge-Based Federated Learning

​

    • The non-convex FEDL problem is decomposed into three convex subproblems and obtains closed-form solutions

CPU-cycle control

Uplink power control

Local accuracy control

  • Theorem 1. The globally optimal solution to FEDL is the combined solutions to three sub-problems SUB1, SUB2, and SUB3.

Solution Approach

[1] Nguyen H. Tran, Wei Bao, Albert Zomaya , Minh N.H. Nguyen and Choong Seon Hong, “Federated Learning over Wireless Networks: Optimization Model Design and Analysis,” IEEE International Conference on Computer Communications (INFOCOM 2019)

37/62

NETWORKING

INTELLIGENCE LAB

38 of 62

Edge-Based Federated Learning

[1] Minh N. H. Nguyen, Nguyen H. Tran, Yan Kyaw Tun, Zhu Han, Choong Seon Hong, “Toward Multiple Federated Learning Services Resource Sharing in Mobile Edge Networks,” IEEE Transactions on Mobile Computing, DOI: 10.1109/TMC.2021.3085979

  • We formulate the multiple FL services Resource Sharing problem (i.e., MS-FEDL Problem) and solve the problem using the decentralized optimization algorithm (i.e., JP-ADMM)
  • The proposed algorithm allows each learning service to independently manage the local resource and learning process without revealing the learning service information

Multiple Federated Learning Services Resource Sharing in Mobile Edge Networks

38/62

NETWORKING

INTELLIGENCE LAB

39 of 62

Edge-Based Federated Learning

  • Support vector machine (SVM)-based federated learning (FL) algorithm method enables each HAB to cooperatively build an SVM model for proactive user associations.
    • Without any transmissions of historical user association results nor of the data size of the task requested

​

  • Given the association decision, the service sequence and task allocation of each user can be optimized to minimize the weighted sum of the energy and time consumption.

Federated Learning for Task and Resource Allocation in Wireless High-Altitude Balloon (HAB) Networks

S. Wang, M. Chen, C. Yin, W. Saad, C. S. Hong, S. Cui, and H. V. Poor, "Federated Learning for Task and Resource Allocation in Wireless High-Altitude Balloon Networks," in IEEE Internet of Things Journal, vol. 8, no. 24, pp. 17460-17475, Dec.15, 2021

39/62

NETWORKING

INTELLIGENCE LAB

40 of 62

Democratized Learning

    • Democratized Machine Learning Architecture is a distributed learning system with a hierarchical structure.
    • The learning model adopted by a large number of clients participate has better learning performance than the learning model derived from the participation of a relatively small number of clients.
    • It is a structure that applies the majority rule adopted in a democratic society.
    • The higher-level and larger groups have more capabilities to solve complex problems via the collective contributions of their members.

​

​

Democratized Machine Learning Architecture

40/62

NETWORKING

INTELLIGENCE LAB

41 of 62

Democratized Learning

​

​

  • Unique features of the democracy in future distributed learning systems
    • According to the difference in characteristics, the learning agent forms an appropriate group that can be specialized so that similar agents can handle the learning task, thereby reducing individual bias through shared generalized learning knowledge and improving learning performance

​

    • Learning agents are free to join, quit, move to any of the appropriate groups and exhibit equal power in the construction of their groups’ generalized learning model

​

    • The power of each group can be represented by the number of its members which varies over the training time

​

​

​

​

​

  • a

M. N. H. Nguyen, S. R. Pandey, K. Thar, N. H. Tran, M. Chen, W. Saad, and C. S. Hong, “Distributed and democratized learning: Philosophy and research challenges,” IEEE Computational Intelligence Magazine 16.1 (2021): 49-62.

Democratized Machine Learning Architecture

41/62

NETWORKING

INTELLIGENCE LAB

42 of 62

Democratized Learning

    • Recent neuroscience research studies the continual life-long learning capabilities for a general artificial intelligence as in biological intelligence
      • Generalization capabilities due to high synaptic plasticity level allow it easier to adapt and learn new knowledge
      • Specialization capabilities increase the specific complex skills
    • The duality of processes in distributed learning
      • The generalized process
        • The high-level of plasticity, the easier to change the group members
        • Generalization broadens the knowledge by sharing among members
      • The specialized process
        • Specialized learning exploits the personalized data at learning agents
        • Encourage a separation of groups due to the personalized characteristics => Groups become stable.

​

​

​

Democratized Machine Learning Architecture

42/62

NETWORKING

INTELLIGENCE LAB

43 of 62

Democratized Learning

  • Democratized Learning (Dem-AI in short) focuses on the study of dual (coupled and working together) specialized-generalized processes in a self-organizing hierarchical structure of large-scale distributed learning systems.
  • The specialized and generalized processes operate jointly towards an ultimate learning goal identified as performing collective learning from biased learning agents.
  • Forces of stability and plasticity are structured to operate according to the learning requirements around the axis of balance.

​

  • a
  • Specialized Process
  • Generalized Process
  • Hierarchical structuring

Fig.1: Conceptual architecture of the democratized learning philosophy

M. N. H. Nguyen, S. R. Pandey, K. Thar, N. H. Tran, M. Chen, W. Saad, and C. S. Hong, “Distributed and democratized learning: Philosophy and research challenges,” IEEE Computational Intelligence Magazine 16.1 (2021): 49-62.

Conceptual Architecture of the Democratized Learning Philosophy

43/62

NETWORKING

INTELLIGENCE LAB

44 of 62

Democratized Learning

Key mechanisms in Dem-AI systems

44/62

NETWORKING

INTELLIGENCE LAB

45 of 62

The Operation of Democratized Learning System

  • The self-organizing hierarchical architecture of the Dem-AI system evolves to adapt to the training environment.

​

  • a

The operation of Dem-AI systems

45/62

NETWORKING

INTELLIGENCE LAB

46 of 62

Hierarchical Generalization and Learning Mechanism

    • The K level hierarchical structure emerges via agglomerative clustering
    • We derive the hierarchical learning problems and there exists the coupling between the upper and lower levels
    • Hierarchical generalized learning problem (HGLP)  for each group i at level k:

​

​

​

​

where Ng(K) is the total number of learning agents in the system,

 Ng,j(k-1) is the number of learning agents of the subgroup j at level k-1,

μ is the trade-off parameters to control the impact of upper group or lower-level subgroups

​

    • The subgroups which have more learning agents have higher impact to the generalized model at level k

​

​

Minh N. H. Nguyen, Shashi Raj Pandey, Tri Nguyen Dang, Eui-Nam Huh, Nguyen H. Tran, Walid Saad, Choong Seon Hong, “Self-organizing democratized learning: Towards large-scale distributed learning systems,” Early Access in IEEE Transactions on Neural Networks and Learning Systems

Close to subgroups models

Close to upper group model

46/62

NETWORKING

INTELLIGENCE LAB

47 of 62

Hierarchical Generalization and Learning Mechanism

    • The closed form of the optimal solution for HGLP can be handily derived by setting the gradient of the objective function to zero as follows:

​

​

​

​

    • The learning model of group i at level k can be updated as

​

​

where

​

w(0)

w(1)

w(2)

w(K)

w(1)t

w(0)t+1

w(1)t+1

1

2

3

w(K)t+1

4

5

6

w(K-1)t+1

w(2)t+1

Upward (bottom -> top)

Downward update (top -> bottom)

7

w(1)t+1

The operation of hierarchical update

for learning models in DemLearn algorithm

Minh N. H. Nguyen, Shashi Raj Pandey, Tri Nguyen Dang, Eui-Nam Huh, Nguyen H. Tran, Walid Saad, Choong Seon Hong, “Self-organizing democratized learning: Towards large-scale distributed learning systems,” Early Access in IEEE Transactions on Neural Networks and Learning Systems

47/62

NETWORKING

INTELLIGENCE LAB

48 of 62

Democratized Learning Algorithm

Personalized Learning

Hierarchical Learning Model Update

Minh N. H. Nguyen, Shashi Raj Pandey, Tri Nguyen Dang, Eui-Nam Huh, Nguyen H. Tran, Walid Saad, Choong Seon Hong, “Self-organizing democratized learning: Towards large-scale distributed learning systems,” Early Access in IEEE Transactions on Neural Networks and Learning Systems

48/62

NETWORKING

INTELLIGENCE LAB

49 of 62

Democratized Learning Algorithm

  • Experiment Settings:
    • We validate the efficacy of the DemLearn algorithm with the MNIST, Fashion-MNIST, Federated Extended MNIST, and CIFAR-10 datasets with 50 clients.
    • The Python implementation of our proposed algorithm using Pytorch and datasets are available at: https://github.com/nhatminh/Dem-AI.
  • Evaluation Metrics:
    • We conduct evaluations for specialization (C-SPE) and generalization (C-GEN) of learning models at agents on average that are defined as the performance in their local test data only, and the collective test data from all agents, respectively.
      • Global : Global model performance
      • G-GEN : Average generalization of group models on average
      • G-SPE : Average specialization performance of group models
      • C-GEN : Generalization performance of learning models at agents
      • C-SPE : Specialization performance of learning models at agents

​

​

​

Minh N. H. Nguyen, Shashi Raj Pandey, Tri Nguyen Dang, Eui-Nam Huh, Nguyen H. Tran, Walid Saad, Choong Seon Hong, “Self-organizing democratized learning: Towards large-scale distributed learning systems,” Early Access in IEEE Transactions on Neural Networks and Learning Systems

49/62

NETWORKING

INTELLIGENCE LAB

50 of 62

DemLearn Algorithm : Experimental Results

Minh N. H. Nguyen, Shashi Raj Pandey, Tri Nguyen Dang, Eui-Nam Huh, Nguyen H. Tran, Walid Saad, Choong Seon Hong, “Self-organizing democratized learning: Towards large-scale distributed learning systems,” Early Access in IEEE Transactions on Neural Networks and Learning Systems

50/62

NETWORKING

INTELLIGENCE LAB

51 of 62

DemLearn Algorithm : Experimental Results

Federated Extended MNIST dataset.

CIFAR-10 dataset

Minh N. H. Nguyen, Shashi Raj Pandey, Tri Nguyen Dang, Eui-Nam Huh, Nguyen H. Tran, Walid Saad, Choong Seon Hong, “Self-organizing democratized learning: Towards large-scale distributed learning systems,” Early Access in IEEE Transactions on Neural Networks and Learning Systems

51/62

NETWORKING

INTELLIGENCE LAB

52 of 62

Edge-assisted Democratized Learning

Shashi Raj Pandey, Minh N. H. Nguyen, Tri Nguyen Dang, Kyi Thar, Nguyen H. Tran, Zhu Han, Choong Seon Hong, “Edge-assisted Democratized Learning Towards Federated Analytics,” IEEE Journal of Things (Early Access) June 2021.

52/62

NETWORKING

INTELLIGENCE LAB

53 of 62

Edge-assisted Democratized Learning : Experimental Results

  • Implementation of DemLearn algorithm is available at

https://github.com/nhatminh/Dem-AI/

  • Developping the device association and resource allocation for this system

Regional Dem-AI learning performance

Shashi Raj Pandey, Minh N. H. Nguyen, Tri Nguyen Dang, Kyi Thar, Nguyen H. Tran, Zhu Han, Choong Seon Hong, “Edge-assisted Democratized Learning Towards Federated Analytics,” IEEE Internet of Things Journal, vol. 9, no. 1, pp. 572-588, 1 Jan.1, 2022

53/62

NETWORKING

INTELLIGENCE LAB

54 of 62

Future Direction on AI for Networking

54/62

NETWORKING

INTELLIGENCE LAB

55 of 62

Future Direction on AI for Networking

  • Designing an explainable multi-modal reinforcement artificial intelligence framework
    • For capturing the dynamics of uncertain environment by relying on beyond the distribution
      • Evidence and rule-based system design
    • By combining neuro-symbolic approach so that agents can trust each other
      • Knowledge graph-based analytics
      • More exploitation than the exploration
  • Network and computational service resource management
    • Industrial IoT and smart production chain
    • Connected and autonomous vehicle systems
    • Smart grid infrastructure
    • Healthcare and cyber physical system

​

How to characterize the trustworthiness among the multi-modal agents for networking?

55/62

NETWORKING

INTELLIGENCE LAB

56 of 62

Future Direction on AI for Networking

Distributed Edge AI for Networking

Challenges of FL/DL over Wireless Networks: poor model quality, resources management, large training time

Statistical heterogeneity

System-level heterogeneity

Communication bottlenecks

Privacy concerns

Challenges of FL/DL over Wireless Networks

Clients selection

  • non-i.i.d. data
  • unbalanced dataset
  • personalized data

Algorithmic design

  • limited wireless resources
  • intermittent connectivity
  • dynamic channel conditions
  • stragglers
  • free-riding problem
  • adversary nodes
  • long-term stay
  • hardware capabilities
  • computing power
  • storage/memory
  • exposed local parameters
  • adversary nodes
  • compromised aggregator
  • convergence time
  • model size
  • network topology
  • aggregation methods
  • computation-communication methods
  • optimization

56/62

NETWORKING

INTELLIGENCE LAB

57 of 62

Future Direction on AI for Networking

    • Develop a novel algorithm design for a multi-task distributed learning setting
    • Integrate advanced multi-level aggregation such as self-attention mechanism
    • Enhancement of privacy and security issues in distributed learning systems:
      • Information exploitation: reverse the personal data
      • Free-riding
      • Model/data poisoning attacks

​

    • Optimization design regarding the synergy of Resource Allocation and Learning Performance
      • Group structure changing

​

    • Future Personalized Applications
      • Learn the unique features and personalized characteristics during the daily activities of each user and make appropriate decisions: Life-long continuous learning
      • VR/AR, and Metaverse services: regional edge intelligence is used to predict the future gaze direction, motion, and mobility patterns, which are exceedingly different among users.

​

Democratized Learning for Networking

57/62

NETWORKING

INTELLIGENCE LAB

58 of 62

Future Direction on AI for Networking

AI for Digital Twin

L. U. Khan, W. Saad, D. Niyato, Z. Han and C. S. Hong, "Digital-Twin-Enabled 6G: Vision, Architectural Trends, and Future Directions," in IEEE Communications Magazine, vol. 60, no. 1, pp. 74-80, January 2022

58/62

NETWORKING

INTELLIGENCE LAB

59 of 62

Future Direction on AI for Networking

    • Semantic networking focuses on the effective transfer of semantic information
      • Extracted key information from a huge data, as opposed to the precise reception of each symbol or bit independent
    • The semanticity in networking can be viewed from different angles
        • A networking model between two devices that can effectively pass information that is meaningful to both source and the destination without wasting a huge bandwidth
        • It can be viewed as effective knowledge gain by an agent with less training from the knowledge of the previous agents from their environment, etc.
    • In a wireless networking environment,
      • Different agents can have unique abilities and missions
      • Where the operation of the collective (or common) set of agents can be based on similar Deep Reinforcement Learning (DRL) tasks

Semantic AI Networking

59/62

NETWORKING

INTELLIGENCE LAB

60 of 62

Future Direction on AI for Networking

Source: https://irtf.org/

The Internet Research Task Force (IRTF) Active Research Groups (14)

Crypto Forum (CFRG)

Global Access to the Internet for All (GAIA)

Quantum Internet Research Group (QIRG)

Human Rights Protocol Considerations (HRPC)

Measurement and Analysis for Protocols (MAPRG)

Path Aware Networking RG (PANRG)

Thing-to-Thing (T2TRG)

Internet Congestion Control (ICCRG)

Information-Centric Networking Research Group (ICNRG)

Coding for Efficient NetWork Communications (NWCRG)

Computing in the Network Research Group (COINRG)

Decentralized Internet Infrastructure Research Group (DINRG)

Network Management Research Group (NMRG)

Privacy Enhancements and Assessments (PEARG)

60/62

NETWORKING

INTELLIGENCE LAB

61 of 62

Future Direction on AI for Networking

    • Decentralize AI for internet of everything infrastructure
      • Standard protocols design for distributed learning system for networking
      • AI data flow design for digital twin of internet
      • Standard metric development for trustworthy AI of internet services and applications
      • AI based semantic networking protocol design for future internet
      • Deployment and operation of AI systems in next-generation networking infrastructure/testbed
      • Scabble AI system protocol design for trustworthy networking and computation

​

Potential AI Topics for Networking Activities with IRTF

61/62

NETWORKING

INTELLIGENCE LAB

62 of 62

Thank You!

Q/A

​

62/62

NETWORKING

INTELLIGENCE LAB