ACCESS IC LAB
Graduate Institute of Electronics Engineering, NTU
Machine Learning Assisted Intelligent RX Design
機器學習輔助之智慧接收機設計
台大電子所吳安宇教授
ACCESS IC LAB Graduate Institute of Electronics Engineering, NTU
P2
Evolution of Communication Systems
Analog voice
was amazing,
but limited
Mobile voice was amazing,
but consumers
wanted more Mobile
for data
increase
1G established
foundation of mobile
connectivity
introducing mobile
services
[1]
2G digital wireless
technologies
increased voice
capacity delivering
mobile to the
masses
3G optimized mobile
for data enabling
mobile broadband
services, and is
evolving for faster
and better
connectivity
4G LTE delivers
more data capacity
for faster and better
broadband
experiences, and is
also expending in to
new frontiers
ACCESS IC LAB Graduate Institute of Electronics Engineering, NTU
P3
ITU-R Defines 5G Application Scenarios
Enhanced Mobile
Broadband
(eMBB)
Massive Machine Type
Communications
(mMTC)
Ultra-Reliability and
Low-Latency Communications (uRLLC)
Source: Recommendation ITU-R M.2083 on IMT-2020 Vision [2]
ACCESS IC LAB Graduate Institute of Electronics Engineering, NTU
P4
5G Communication Systems
Flexible and diverse requirements in
5G communications
Requirements
High throughput/capacity
Low latency
High reliability
Massive connectivity
Huge gap between 4G and 5G
100x peak data rate [Gbps]
0.1x latency [ms]
10x simultaneous connection [M/km2]
Need a powerful technique to overcome the HUGE gap from 4G to 5G !
[2]
ACCESS IC LAB Graduate Institute of Electronics Engineering, NTU
P5
Intelligent Communication Systems
Communications with the ability of learning = “Cognitive radio” (CR)
J. Mitola III - The “Godfather” of cognitive radio technologies
Dynamic spectrum access
RF sensing – sense the spectrum
Reconfigurable
Radio systems could achieve rapid adaptation, self-optimization, self-
organization, and autonomous operation
Cognitive
radio
Mobile frequencies crowded With CR you can use any channel
[3][4]
ACCESS IC LAB Graduate Institute of Electronics Engineering, NTU
P6
Development of Machine Learning
1950’s
1960’s
1970’s
1980’s
1990’s
2000’s
2010’s
~
Now
Artificial Intelligence (AI)
Any technique which enables
computers to mimic human behavior
Machine Learning (ML)
Subset of AI techniques which
use statistical methods to enable
machines to improve with
experiences
Deep Learning (DL)
Subset of ML which uses neural
network models to understand
large amounts of data
Ex: computer vision, AI gaming,
image & speech recognition…
ACCESS IC LAB Graduate Institute of Electronics Engineering, NTU
P7
Deep Learning (DL)
Heavily investing by Google, Microsoft, Amazon, and other tech. giants
Most seen in image/speech recognition, regression, dimension reduction
Deep learning in 5G has picked up momentum
Self-organizing resource allocation
Network management/routing
…, and so on.
[3]
ACCESS IC LAB Graduate Institute of Electronics Engineering, NTU
P8
Convolutional Neural Network (CNN)
Recurrent Neural Network (RNN)
What is Deep Learning?
Training: define network structure as
neural network
Multiple layers cascading of
nonlinear processing units for
feature extraction and
transformation
Learn high levels of feature /
representations
Automatically features collection
For Sequential Problem
For Computer Vision Tasks
Capture different levels
of feature representation
Memorize and recall the memory
[5]
ACCESS IC LAB Graduate Institute of Electronics Engineering, NTU
P9
Deep Learning + 5G Communications?
• Feature learning by DL itself
• Continuous improving performance
Deep Learning
• Massive connected devices
• Huge amount of communication data
5G Communications
Goal of DL for 5G: Meet requirements and make up the huge gap from 4G to 5G !
Replace the vast expert knowledge in the field of communications
Inadequate system models
Mathematically convenient models - not for real systems with imperfections
Limiting functional block-structure
Individual blocks optimization ≠ the best possible end-to-end performance
Handcrafted feature extraction
Machine can learn high-level feature autonomously which is more representative
Unlimited development potential
[6][7]
ACCESS IC LAB Graduate Institute of Electronics Engineering, NTU
P10
Proposed Intelligent Transceiver Design
(A) Pre-distortion & Equalizer
(B) Automatic Modulation Classification
(C) Neural Network Demodulator & Decoder
(D) End-to-End Optimization
ACCESS IC LAB Graduate Institute of Electronics Engineering, NTU
P11
Exploit deep learning to meet the requirements and make up the huge
gap from 4G to 5G
(A) Pre-distortion & Equalizer
DL-based nonlinear equalizer to eliminate ISI & PAPR low complexity
(B) Automatic Modulation Classification
Dynamically adjust the transmission data rate and avoid handshaking
spectral efficiency
(C) Neural Network Demodulator & Decoder
Avoid data dependency low latency
(D) End-to-End Optimization
Break functional blocks higher system reliability
Conclusion
Machine learning communication systems hold the potential to
improve the capacity, reliability, latency, and energy efficiency.
ACCESS IC LAB Graduate Institute of Electronics Engineering, NTU
P12
Reference
[1] The Evolution of Mobile Technologies: 1G 2G 3G 4G LTE - Qualcomm
[2] Recommendation ITU-R M.2083 on IMT-2020 Vision
[3] Jiang, Chunxiao, et al. “Machine learning paradigms for next-generation wireless networks,” IEEE Wireless
Communications 24.2 (2017): 98-105.
[4] Mendis, Gihan J., Jin Wei, and Arjuna Madanayake, “Deep learning-based automated modulation classification for
cognitive radio,” 2016 IEEE International Conference on Communication Systems (ICCS).
[5] 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).
[6] http://machinelearningmastery.com/improve-deep-learning-performance/
[7] O'Shea, Timothy J., and Jakob Hoydis. “An introduction to machine learning communications systems,” arXiv preprint
arXiv:1702.00832 (2017).