oulu webinar - 6g channel
TRANSCRIPT
Machine Learning in 6G Systems:
An Overview
Oulu Webinar
Walid SaadElectrical and Computer Engineering Department,
Network sciEnce, Wireless, and Security (NEWS) Group
Virginia Tech
Email: [email protected]
Group: http://www.netsciwis.com
Personal: http://resume.walid-saad.com
2
Why ML in 6G?
Why do we need ML in 6G?
Communication and autonomy are intertwined – ML/AI will
already be there!
Communication must be self-sustaining – system should be able
operate itself by itself: less rigid protocols!
Communication should be proactive – learn the user (beyond
basic caching) or learn the comm. medium!
Three use cases for AI/ML
Data analytics for proactiveness (started in 5G and continues)
Artificially-intelligent optimization and control (theoretical in
5G+ and can potentially replace standardized protocols in 6G)
Edge learning and small distributed data (6G)
3
Use Case 1: Data Analytics
Large-scale wireless networks or…
…………data hurricane!
Massive volumes of data will be generated from sensors,
wearables, surveillance cameras, etc.
How to make sense of this hurricane of chaos?
Analyze, predict, and adapt to the wireless user behavior,
user-centric wireless (e.g., virtual reality)
4
Use Case 2: Autonomy and AI-enabled Optimization
Connected devices (and people) will make use of big data
to act, react, and adapt intelligently
ML/AI are not mere data analytics tools
They will instill intelligence into the wireless network design itself:
AI-driven control and optimization
The idea of SON will no longer be simple optimization, it becomes AI-
driven self-sustaining networks!
AI-enabled devices will be there, let’s use them!
Data may or may not be necessary
Intelligent optimization can be trained numerically in an online
manner, but of course data can enhance performance
Reinforcement learning is the key, let’s see a detailed example
5
Reliability and Low Latency
Communication with connected autonomy
Massive number of devices with services that are
more critical than Facebook/Instagram posts!
6
Communication Requirements
How to effectively connect all those infrastructure and
components of a city?
Communication should be high data rate – terrabytes of
data to be delivered!
Communication should be high-speed – low delay and
response time (few millisecond target!)
Communication should be pervasive – anytime,
anywhere, anyone!
Communication should be reliable – seven 9’s for 6G!
Reliable low latency communications a key in 6G
Beyond ultra reliable low latency communication (URLLC)
7
URLLC: What’s next?
URLLC has been around for a while but prior art…
Focused on IoT sensors (uplink) – autonomous vehicles/drones
are different (downlink?)!
Assumes known models for traffic (M/M/1 etc.)– latency has many
components, hard to model!
Considers slow deep reinforcement learning (DRL) – learning in
URLLC must handle extreme, rare conditions (6G feature)!
Assumes rate can be ignored – 6G applications may need some
form of rate guarantees!
Think of a new service, we can call it highly reliable low
latency communications (HRLLC)
Experienced DRL for Model-free HRLLC
8
System Model
Consider the downlink of a single-cell wireless network
whose base station is sending latency-sensitive control
message to autonomous vehicles
We consider a downlink OFDMA system with resource blocks that must be allocated
Reliability is defined as the probability of end-to-end
packet delay exceeding a threshold
We do not make any assumptions for a delay model
Delay is intrinsically hard to model, most models are often
unrealistic and have some hidden drawbacks
Delay has many components, hard to model their combination
precisely
Our goal is to solve the following problem
Explicit rate guarantees imposed
Challenging to solve because of our model-free assumption
Problem Formulation
9
Reliability
constraint
Feasibility
constraints
Rate
constraint
Total power
10
Handling Model-Free
A wireless network can empirically measure the delay
Network can “learn” the delay once it connects with a user
How? Use case 2: Reinforcement learning is natural but…
….classical solutions cannot handle the large state space
Solution: Deep reinforcement learning
Deep RL used because it is appropriate to handle our large
state space not because it is “fashionable”
Ratio of number of packets with delay excess
and total number of packets
11
Deep-RL for Model-Free HRLLC
State space: number of packets transmitted, packet size, and
channel gains
PPO: Proximal policy optimization determines target rates
Action space reducer: Deep-RL made tractable
12
Deep-RL for Model-Free HRLLC
The reward function used by deep-RL:
Theorem 1: By maximizing this reward, after convergence of
the deep-RL algorithm, the reliability of each user is
guaranteed, such that:
But, is deep-RL reliable and suitable for HRLLC?
No! Can be slow to converge and unreliable extreme cases
Solution? Use generative adversarial networks (GANs)!
Time-varying weight
that control
the reliability
13
Experienced Deep RL
Use GAN to create a “virtual environment” for training
Virtual environment is created by GAN using a mix of
(limited) real data and synthetic (simulated) data
14
Experienced Deep RL
GAN-based refiner
Proposed by Apple for
computer vision
Inputs
Unlabeled real data
Synthetic model data
Output
Refined (and larger) dataset
that includes new network
conditions (extreme events)
that can train your deep RL
15
Simulation Results
We use a real dataset with specific packet sizes and
interarrival times (with some modification)
16
Simulation Results
Experience allows a very smooth handling of extreme
events compared to vanilla deep-RL
See paper: A. T. Z. Kasgari, W. Saad, M. Mozaffari, and H. V. Poor, "Experienced Deep
Reinforcement Learning with Generative Adversarial Networks (GANs) for Model-Free Ultra Reliable
Low Latency Communication", IEEE Transactions on Communications, to appear, 2020.
Open Questions
Use of a neural network instead of PPO to enhance/reduce
possible overhead
More rigorous approach to action space reduction
Extensions to multi-agent scenario is a very interesting
aspect (how to look at scale in that case)
We used a deep Q network
Can we design new deep RL architectures with a more
interesting backbone ANN? (see M. Chen, W. Saad, et al.)
More control on the GAN?
Useof experienced deep RL in other 6G problems17
Device Device Device Device
Federated AI (current)Collaborative AI (future)Classical AI (past)
• Cloud-AI w/ dumb devices
• All data in the cloud
• Classification/inference at the cloud
• No privacy
• No cloud and/orinfrastructure needed
• Collective intelligence
• Privacy-preserving
• Cloud-AI + on-device AI = mobile AI
• Use the cloud but smartly
• Privacy-preserving
Use Case 3: Edge AI
Massive Small Data and Distributed Machine learning
Cloud big data was the norm but data is private and
distributed
Can users collaboratively
learn a task of interest?
Learning (at the edge) to
communicate?
Communication for
learning (joint design)?
Areas ripe for exploration
in 6G
Joint FL-Wireless Design
20
Since the global and local FL model
parameters need wireless link exchanges,
then wireless transmission errors will
impact FL performance
The base station (BS) must update the
global FL model as it receives all of the
local FL model transmitted from the
users. Hence, transmission delay and
energy consumption must be
considered.
Communication to support FL!
FL training process in a wireless network
System Model
21
Consider a cellular network with one BS and U users that
are engaged in an FL algorithm
Uplink transmissions are used for sending the local FL
model parameters
Downlink transmissions are used for updating the global
model parameters from the BS to the users
M. Chen, Z. Yang, W. Saad, C. Yin, H. V. Poor,
and S. Cui, “A Joint Learning and Communications
Framework for Federated Learning over Wireless
Networks”, IEEE Transactions on Wireless
Communications, to appear, 2020.
System Model
22
Local FL model including transmission errors
where
With packet error rates, the global FL model becomes:
PER over RB n
Packet error rate (PER)
If a user i is
discarded because
of high PER then:
Problem Formulation
23
RB allocation matrix
User association vector
Delay constraint
Energy constraint
Power allocation matrix
Theorem 1: An upper bound on the convergence point
of FL over wireless networks can now be found:
This is a key characterization of FL performance over a
wireless network
Convergence affected by PER and user association: wireless
network must be reliable enough to support effective FL
We used full gradient descent but results extendable
Global FL Model Analysis
24
Total samples
Number of samples
of user iGlobal model at
convergence
FL Optimization
25
Based on Theorem 1, we can solve the original
optimization problem and find the optimal uplink
transmit power of a user i for a given user selection a
and uplink RB allocation vector ri:
where satisfies the equality
We can also find the optimal uplink RB allocation by
using a very simple Hungarian method
This is a different view of learning: We optimize a
wireless network to serve edge FL!
Joint learning and communication design: A new paradigm
Simulation Results
Value of the
loss function
as the number
of samples per
user
varies.
Simulation Results
27
Digit above: Identification result of the proposed FL
Digit below: Identification result of the standard FL
Red digit: Identification error
An example of
implementing
FL algorithms
for handwritten
digit
identification.
Remarks
We now can ask: What can 6G do for FL, rather than the
reverse?
We cannot yet give a direct answer but…
We need reliability and low latency for FL convergence
Data rate may also be needed (some FL models are large!)
Joint learning and communications is hot for beyond 5G
How to maintain high accuracy, low convergence time, and
error-free communications, jointly?
Generative models (brainstorming GANs: A. Ferdowsi and W.
Saad)
Beyond FL: full distributed/collaborative learning 28
29
What’s next in AI?
Issues of reliability (particularly at high frequencies):
Experienced RL is one idea, but are there others
Distributed learning – how accurate can we go?
Beyond standard learning:
Learning is mostly training-based, can we get rid of training?
We are pursuing several ideas in this context: learn one task
and map it to another (like humans do), this includes self-
learning and beyond meta-learning
Theoretical foundations for RL/explainable AI
Topological aspect (persistent homology)
Rich field of application in 6G with specific needs
Acknowledgement Acknowledgement to students/collaborators: Ali Taleb Zadeh Kasgari,
Mingzhe Chen, Tengchan Zeng, Anibal Sanjab, Christina Chaccour,
Omid Semiari, Mohammad Mozaffari, Gilsoo Lee, Mehdi Bennis, and
Merouane Debbah
More on 6G? W. Saad, M. Bennis, and M. Chen, "A Vision of 6G Wireless Systems:
Applications, Trends, Technologies, and Open Research Problems", IEEE Network,
2020 (February 2019 pre-print on Arxiv).
More on AI? M. Chen, U. Challita, W. Saad, C. Yin, and M. Debbah, "Artificial
Neural Networks-Based Machine Learning for Wireless Networks: A Tutorial", IEEE
Communications Surveys and Tutorials, vol. 21, no. 4, pp. 3039-3071, 2019
More on FL? M. Chen, H. V. Poor, W. Saad, and S. Cui, "Wireless
Communications for Collaborative Federated Learning", IEEE Communications
Magazine, to appear, 2020.
More on self-learning? Y. Xiao, G. Shi, Y. Li, W. Saad, and H. V. Poor, "Towards
Self-learning Edge Intelligence in 6G", IEEE Communications Magazine, Special
Issue on Communication Technologies for Efficient Edge Learning, to appear, 2020.30
Finally….
Thank You