Dell's Distinguished Engineer Wenjing Chu discusses innovations in applying Machine Learning to solve challenges in Telco/Communication Services, and predicts that the future is a Sentient Network powered by Machine Learning that can handle real-time high-velocity data.
A Sentient Network: How High-velocity Data and Machine Learning will Shape the Future of Communication Services
1. A Sentient
Network
How High-velocity Data and Machine Learning
will Shape the Future of the Communication Services
OPNFV Summit, Burlingame CA,
November 9-12, 2015
Wenjing Chu
Distinguished Engineer
Dell Research
Member of the Board and TSC of OPNFV
3. “The more real-time and
granular we can get, the more
responsive, and more
competitive, we can be.”
Peter Levine | Andreessen Horowitz
A Sentient Network
1
Elastic on-demand capacity
Open software architecture promises flexible elastic capacity that
can be rapidly provisioned and dynamically managed
Data-driven operation automation
Virtualization unleashes the latent value in the real-time data to
optimize resource allocation and assure SLA
Scalable infrastructure
Standard open architecture infrastructure delivers capacity, cost
efficiency, and right-sized reliability
2
3
Self-learning security and privacy
Self-learning algorithms from real-time data delivers ultimate
security and privacy at the same time
4
Machine intelligent user services
Advances in Machine Learning promise continuous improvements
in user experience
5
4. New Paradigm Shift in Infrastructure: NFV/SDN
Domain specialized software on standard
hardware, delivered from the cloud
- Dramatically cuts CapEx & OpEx
- Enhances service velocity
- Enables Big Data driven business model
5. High-velocity Cloud Empowers Business Transformation
Mobile' Infrastructure
Content' Distribution
Edge Computing
High'Velocity'Cloud
Packet Velocity
• 100X moreperformance
• 50X morecustomers
Service Velocity
• Deploy services in minutes vs months
• Empower new, innovative business
models
Data Analytics Velocity
• Sub-second real-time streaming
analytics
• Sentient intelligence
6. High-velocity Data with Machine Learning
Telemetry,
IoT
sensors,
System
logs,
Monitors,
Mobile
devices …
Transmi
ssion of
data in
streams
Transfo
rmation
Learning
in real-
time
Action
on
intellige
nce
11. Automatically Adjust Resources to Maintain SLA
Systems can respond to usage spikes in real-time,toreallocate resources and
maintainSLAs.
12. Continuous Resource Optimization by Reinforcement
Learning
! Modeled as a Markov
Decision Process
! Learning probability
distributionby Bayesian
inference
! Q-Learning, Deep Q-
Network
! Consensus optimization
Wikipedia: MDP
14. Classification by Concept Adapting Decision Tree
! Rules programmingis
labor intensive,error
prone, static
! Let algorithm learns a
DT (or a forest) on its
own
! Concept adaptability:
incorporate new, forget
old
Packets > 10
yes no
Protocol=http
Packets > 10
no
yes
Bytes > 60k
yes no
Protocol=ftp
Data stream
Data stream
15. Uncovering Unusual Hidden Activity by Monitoring Entropy
! Entropy in a moving
time window captures
the normal hummingof
the system
! Out of ordinarymove of
entropy plus context
suggest attack vs. flash
crowd
16. Clustering Users based on Behavior Patterns
! Non-parametric model
can be used for latent
features,overlapping
clusters and infinite data
! Eg Dirichlet process,
Gaussian process
! A cluster of ‘users’of
abnormal behavior are
suspects
18. ! Mining telco CDR’s to
evaluate risks from
customer churn
! Combining locationand
real-time system infoto
pinpoint qualityissues
! Machine learning
algorithm offers more
precision
Proactive Customer Support and Retention
The peaks indicate areas of highest risk with more
precision than traditional linear regression (the dotted
line).
Creative commons http://scicomp.stackexchange.com/
20. 20
“How is Seamless Mobilitypowered by High
Velocity Cloud?”
Seamless Mobility by Contextual Learning
Live machine learning
algorithms ensure
quality, security and
seamless mobility.
High-velocity Cloud
High-velocity Analytics
22. Differential Privacy in Big Data and Machine Learning
! Anonymizationis not
enough
! Differential Privacy(!-
DP) provides a formal
guarantee & a
mechanism for tradeoff
! DP may also help avoid
False Discovery
Dr.Katrina Ligget, CalTech
23. Computing on and Learning from Encrypted Data
transformed+
queryplain+query+
under+passive+attack
Application
decrypted+
results
encrypted+
results
DB+server
encrypted+DB
Proxy
Secret
Secret
computation+on+
encrypted+data+≈+
regular+computation
! Stores+schema++
and+master+key
! No+query+execution
trusted+client?side
! Data loss is prevalent
everywhere you look
! Data privacy
responsibilityis unclear
! Practical system can be
deployed with strong
encryptionwithout the
risk of key disclosure
! Different algorithm for
different computation
Dr. Laruca Popa, UC Berkeley
24. So, Any Takeaways for OPNFV ?
• Collect data
• Put data in an open format
• Consider privacy and security on day one
• Don’t tie data to a specific implementation of a specific design
• Must consider the time dimension of data, e.g. TSDB, streaming
25. “The future is already here – it’s just not very evenly distributed.”
William Gibson