"Toward Cognitive-IoT Applications -- Integrating AI with Fog Computing" by Dr. Frank C. D. Tsai, Workshop of Mobile IoT with Edge Computing and Artificial Intelligence, sponsored by Ministry of Education, Taiwan
(INDIRA) Call Girl Bhosari Call Now 8617697112 Bhosari Escorts 24x7
20180115 Mobile AIoT Networking-ftsai
1. III Proprietary. All rights reserved. 0
教育部
行動寬頻尖端技術
跨校教學聯盟
• 行動寬頻網路與
應用 -- 行動智慧
聯網聯盟中心計
畫
2. III Proprietary. All rights reserved.
Speaker: Frank Chee-Da Tsai, Ph.D. 蔡其達
R&D Director, Digital Transformation Institute, III
with Prof. Fuchun Lin (NCTU)
Toward Cognitive-IoT Applications
--Integrating AI with Fog Computing
January 15, 2018
1
3. III Proprietary. All rights reserved.
Digital Transformation
Source: The Battle Is For The Customer Interface, Tom Goodwin, Havas Media, March 2015 -
http://techcrunch.com/2015/03/03/in-the-age-of-disintermediation-the-battle-is-all-for-the-customer-interface/
The world’s largest
taxi company…
… owns no cars
The world’s most
popular media
owner…
… creates no
content
The world’s most
valuable retailer…
… owns no stores
The world’s largest
accommodation
provider…
… owns no estate
… … …
… … …
Disruptive business models -- cyber customer interface has become more
important than physical presence
The Key Success Factor for DX ( when impacted by 3rd platform)
• Instant analytics and precise prediction… Cognition (by AI+HI)
• action across systems and organizations… Control (via edge-fog-cloud)
2
4. III Proprietary. All rights reserved.
Artificial Intelligence Timeline
Source: Gartner, 2017
By 2020
• 30% of CIOs will include AI in their top five investment priorities.
• 30% of new development projects will have AI components delivered
by joint teams of data scientists and programmers.
Supervised Learning, Unsupervised Learning, Reinforcement Learning all
benefit from Deep Learning – repetitive layers of Synthesis-Summary
3
5. III Proprietary. All rights reserved.
AI is Driving H/W Evolution
Source: Gartner, 2017
By 2020
• Deep neural networks (DNNs) and machine learning applications will
represent a $10 billion market opportunity for semiconductor vendors.
AI is to gold rush in California in the late 1840s what xPU semiconductor is
to picks and shovels supplied to the miners -- AI application startups may or
may not prosper, but … semiconductors will.
4
Source: http://bangqu.com/65Co2N.html, 2017
6. III Proprietary. All rights reserved.
The global competitive focus has shifted from the Internet of Things (IoT)
to Artificial Intelligence (AI)
ICT industry competition has moved from mere hardware, to hardware
and software integration, to an integrated solution for service innovation
through -- machine learning, natural language processing, speech
recognition, computer vision and other AI technologies.
• From ‘Mobile-First” to “Ai-First” – google’s strategic development direction
Global Competitive Edge
Autonomous intelligence can sense, learn and change the environment
5
7. III Proprietary. All rights reserved.
Anatomy of an IoT System
Source: Cognitive Internet of Things: Making Devices Intelligent, IBM, 2017
Cloud-centric IoT System – IoT Cloud Platform & Services (PaaS & SaaS)
provides IoT applications and solutions.
cognition
6
The capability of description, prediction, prescription toward cognition is
the competitive edge.
8. III Proprietary. All rights reserved.
Cognitive Computing
Source: WikiPedia, 2017
CC mimics the functioning of the human brain and helps to improve
human decision-making. (….. Still have a way to go though.)
Technology platforms -- encompass machine learning, reasoning, natural
language processing, speech recognition and vision (object recognition),
human–computer interaction, dialog and narrative generation, etc.
• Contextual: They may draw on multiple sources of information,
including both structured and unstructured digital information, as well
as sensory inputs (visual, gestural, auditory, or sensor-provided)
• Adaptive: They may be engineered to feed on dynamic data in real
time, or near real time. They may learn as information changes, and as
goals and requirements evolve.
• Interactive: They may interact with other processors, devices, and
Cloud services, as well as with people.
• Iterative and stateful: They may aid in defining a problem by finding
additional source input if a problem statement is ambiguous or
incomplete. They may "remember" previous interactions.
IBM Watson offers many Cognitive APIs (e.g. Natural Language Classifier,
Sentiment Analysis, Relationship Extraction, Visual Recognition, …)
7
9. III Proprietary. All rights reserved.
Cognitive IoT Platform
Cognitive IoT Solution Architecture -- a system with feedback control loop
• Data & Measurement Platform (sensors data capturing)
• Modeling & Orchestration Platform (platform cognition)
• Control Platform (application actuation)
Source: Enterprise Information Systems, Taylor & Francis, Chung-Sheng Li, 2017
1
2
4
3
8
contextual
interactive
interactive
adaptive
10. III Proprietary. All rights reserved.
An Example – Smart Grid
Control-loop continuously optimizes the expected outcome via dynamic
data-driven behavior models
9
1
2
4
3
9
Source: Enterprise Information Systems, Taylor & Francis, Chung-Sheng Li, 2017
11. III Proprietary. All rights reserved.
AI in the Cloud
Google, Amazon, Microsoft and IBM leading the charge
AI (Machine Learning/Deep Learning) PaaS
• Azure Machine Learning, AWS Machine Learning, Google Cloud
Machine Learning
• Enable the creation of machine learning models using a specific
technology on their platform
• Execution of AI programs written in mainstream AI frameworks like
Theano, Torch, TensorFlow, Caffe, etc. mostly not supported
AI (Machine Learning/Deep Learning) SaaS
• IBM Watson, Microsoft Cognitive Services, Google Cloud Vision or
Natural Language APIs
• Allows applications to incorporate (general, standardized) AI capabilities
by using offered AI or cognitive computing APIs, without having to invest
in sophisticated AI infrastructures
source: The AI-First Cloud: Can artificial intelligence power the next generation of cloud computing? By Jesus Rodriguez, 2016
10
12. III Proprietary. All rights reserved.
Cloud APIs for AI-as-a-Service
AI Service
AI Platform
(ML)
AI Engine
(DL)
AI Basic
Framework
Information
Manageme
nt
Big Data
Storage
Machine
Learning
Business Model – mostly horizontal applications, priced by mixed package of
number of uses, time of use and storage usage.
source: IEK, 2017
11
13. III Proprietary. All rights reserved.
Pure Cloud Service Model
In a pure cloud-centric model, all raw data are aggregated and streamed to
the cloud for storage and processing.
Despite the advantages, the model has some major drawbacks:
• Unreliable cloud connections can bring down the service
• Unpredictable response time from cloud server to endpoints
• Excessive data can overburden infrastructure
• Privacy issues when sensitive customer data are stored in the cloud
• Difficulties in scaling to ever increasing number of sensors and actuators
The drawbacks of cloud-centric design makes it unsuitable for mission critical
industrial such as manufacturing, healthcare, transportation applications.
Source: Intelligent IoT and Fog Computing Trends, By Frank Lee -September 14, 2017
12
14. III Proprietary. All rights reserved.
Back to the Future
Centralized v.s. Distributed Computing -- Those long divided shall be united;
those long united shall be divided: such is the way of the universe.
Source: 2017 Asia Fog Summit, Junshan Zhang , 2017/12
13
15. III Proprietary. All rights reserved.
Cloud to the Edge
:
What is old is new again
Source: Gartner, 2017
14
16. III Proprietary. All rights reserved.
Cloud to the Edge
What is old is new again
15
Source: Gartner, 2017
17. III Proprietary. All rights reserved.
Cloud to the Edge
What is old is new again
16
Source: Gartner, 2017
18. III Proprietary. All rights reserved. 17
What is old is new again
Source: Gartner, 2017
Cloud to the Edge
19. III Proprietary. All rights reserved.
Cloud to the Edge
What is old is new again
18
Source: Gartner, 2017
20. III Proprietary. All rights reserved.
Cloud to the Edge
What is old is new again
19
Source: Gartner, 2017
21. III Proprietary. All rights reserved.
Industry Are Moving to the Edge
MEC
2014
Standards for mobile
edge computing
2015
Turning wired telecom
central switches into local
data centers
Cloudlet
2009 2015
Developing ways to miniaturize clouds
Multiple Consortia Are Developing Silo-ed Solutions for Different Market Verticals
HPE Edgeline Gateways
(Edge computing, data capture,
analytics, and control)
Dell Edge Gateways
(Edge analytics: aggregate,
analyze, filter data)
Edge computing clusters
for SP mobile networks
Mobile Edge Cloud,
Mini Clouds
Major Players Have Been Developing and Selling Fog-like Products
GE Predix + Microsoft Azure
for
edge-to-cloud service continuum
Edge Data Center
Content Delivery Networks
IOx,
Fog Director, Fog Data
Services, …
Oracle Edge Applications
Source: Cisco, Tao Zhang , 2017
20
22. III Proprietary. All rights reserved.
What’s Wrong with the Picture?
Each industry is developing its own solutions, resulting in silo-ed solutions for
• Different networks: 5G, wired telecom, enterprises
• Different industry verticals: manufacturing, smart cities, ...
• Different applications inside the same industry verticals: traffic signal
control, traffic surveillance, lighting control, …
Isolated edge systems and applications
• Not integrated or poorly integrated with the cloud
• Difficult to interoperate or collaborate with each other
Market and Customers massively confused
• Edge Computing vs. Mobile Edge Computing vs. Multi-access Edge
Computing vs. Mobile Edge Cloud vs. Cloud RAN vs. MiniCloud vs.
Cloudlet vs. CORD vs. ...
• and where does the Cloud fit in all these?
21
Source: Cisco, Tao Zhang , 2017
23. III Proprietary. All rights reserved.
Fog as a System-Level Architecture
Cloud Cloud Cloud
E2E Architecture
• Distribute, use, manage, and
secure resources & services
• Enable horizontal and vertical
interoperability, composability,
and automation
(not just placing servers, apps,
or small clouds at edges)
Cloud-to-Thing Continuum
• Enable computing anywhere
along the continuum
(not just at any specific edge)
• Orchestrate resources in
clouds, fogs, and things
(not just isolated edge devices,
systems, or apps)
Horizontal
• Support multiple network types
and industry/application
verticals
(not silo-ed systems for
different network types,
industries, or application
domains)
Works Over and Inside
• wireless and wireline networks
(no need for separate
platforms just for moving
computing inside any specific
type of network such as 5G)
Fog
Horizontal system-level
architecture that
distributes computing,
storage, control, and
networking functions
closer to users along the
cloud-to-thing continuum
22
Source: Cisco, Tao Zhang , 2017
24. III Proprietary. All rights reserved.
Fog Computing as Common
Distributed Computing Framework
Fog Computing Is Analogous to Previous Internet Revolutions: TCP/IP, WWW, …
Wouldn’t it be better if we
developed a TCP/IP-for-wireless
telecom? a TCP/IP-for-wired? a
TCP/IP-for-enterprise? … NO
TCP/IP
A horizontal framework
for
distributing packets
Fog Computing
A horizontal framework
for
distributing computing functions
and
making use, managing, and securing
these distributed resources and services
Will it be a good idea to develop a
fog-like system for 5G? another for
wired telecom? another for
enterprises? another for smart city?
another for manufacturing? … NO
WWW
A horizontal framework
for
accessing files anywhere
Wouldn’t we be better off with a
HTTP-for-wireless? a HTTP-for-
wired? a HTTP-for-enterprise? … NO
23
Source: Cisco, Tao Zhang , 2017
25. III Proprietary. All rights reserved.
Distributed Computing Frameworks
Characteristics Fog MEC Cloudlet
Networks Datacom + Telecom Telecom Datacom
Usage Scenarios Volumable IoT 5G Resource-limited
Mobile devices
Computing location Fog Nodes & Network Edge Servers Cloudlets
Network Structure Internet + Cellular Cellular Cellular
Corresponding Eq. Various IoT Equipment Mobile devices,
Vehicles
Mobile devices (e.g.
handsets)
Main Standard body IEEE/NIST ETSI NIST
Service Types Public, Enterprise Public Enterprise
Network Topology Hierarchical +
Horizontal structure
Hierarchical
structure
Hierarchical
structure
Key Techniques Cloud-Fog
Orchestration, Flexible
deployment of Fog
nodes to adapt for
different service needs
Convergence of
Edge server and
base stations,
hierarchical
deploment
Small cloud servers
to offer local
computing resources
and applications
Source: MIC, 2017
24
26. III Proprietary. All rights reserved.
From Cloud-centric to Node-based
Cloud (as well as cognition APIs in the Cloud) is now descending to be
diffused among the client devices (so is congnition/AI in the device), which
are often with mobility too.
The cloud is becoming “fog,”
where Fog Nodes --
• form a mesh to provide
resilience, fault tolerance,
load balancing, with min.
cloud communication.
• communicate laterally
(peer to peer) and
communicate up and
down with IoT and Cloud.
• are able to discover, trust,
and utilize the services
of another node in order
Source: IEEE-SA, PAR1934, 2017
to sustain RAS(reliability, availability, serviceability)
25
27. III Proprietary. All rights reserved.
Hybrid Fog/Cloud Model
Cloud & Fog are to co-exist and cooperate in offering cognitive services to
IoT Applications
Cloud nodes
• handle non-real-time or soft real-time functions
– software update, contextual information collection, long term big data
analysis, ….
Fog nodes
• handle time sensitive AI inference tasks or privacy sensitive pre-processing
• perform real-time AI inference, using data from a large number of sensors
• carry out actions by sending commands to actuators in machines, drones,
or robots
• collect the real-time feedback results, to evaluate the next actions to take or
pass to the proximal fog node for further processing
26
28. III Proprietary. All rights reserved.
AI in Fog Computing
Support for mainstream AI frameworks
• TensorFlow, Caffe, Theano, Torch, etc
AI-first services
• AI being a key element to improve the intelligence of fog node services
such as storage, compute, or security
• Built-in into off-the-shelf equipment as fog nodes
Integration with PaaS/IaaS services in Fog Node
• provide seamless integration between AI and deep learning
frameworks and the existing services framework included in fog
platforms
Management tools
• operational management tools to manage and operate AI modules
deployed in the infrastructure
27
29. III Proprietary. All rights reserved.
Micro SaaS + PaaS + IaaS
28
FaaS (micro-SaaS) + (stripped down) PaaS, IaaS in a single Fog (Edge) Node
Virtualization is to a Fog Node (via Fog PaaS/IaaS) what virtualization is to Cloud
flexible combination to adapt to various services made easy (like Legos)
OpenFog Architecture. v2
28
30. III Proprietary. All rights reserved.
Sample Fog Use Cases
Source: IEEE-SA, PAR1934, 2017
29
Many sub-systems A few bandwidth-eating devcies Lots of, repeated smalll data,
31. III Proprietary. All rights reserved.
Analytic Module in Whole (1)
-- using video surveillance as an example
Gender/Age
Trajectory
BI
Physical Architecture
• intermediate hops act as pure “relay”
• all processing (analytics) done in the cloud
Logical (Software) Architecture
• Analytics using a big “analytical network” – a combination of a series of DL
(series/parallel connection). All processing (analytics) done in the cloud
People
detector
Face
detector
Common
features
extraction
Age Classifier
Gender
Classifier
Tracking
Age
Gender
Trajectory
BI
SSD
SqueezeNet
GoogleNet
OpenCV
Tiem+Space
ColorGradient
30
Source: III Innovation Project, Frank C. Tsai, 2017
32. III Proprietary. All rights reserved.
Breaking Analytic modules into “micro-services’
• So that “early exit” is possible to allow SI/Biz. Op. to save bandwidth
• So that response/actuation is immediate without delay
People
detector
Face
detector
Common
features
extraction
Tracking
SSD
Age
Classifier
Gender
Classifier
Age
Gender
BI
SqueezeNet
GoogleNet
OpenCV
Time+Space
ColorGradient
Possible early exit
If no moving objects into
restricted area at all
Possible early exit
If object do not stay
but leave or retreat
Analytic Module in Part (2)
-- using video surveillance as an example
Mapping Micro-services to Fog Nodes
• So that flexible combination to adapt to various services is made easy
• Sharing services between nodes to achieve the final goal
• Careful to re-use trained deep learning models without re-loading
31
Immediate
response
Source: III Innovation Project, Frank C. Tsai, 2017
33. III Proprietary. All rights reserved.
ILSVRC -- ImageNet Large Scale
Visual Recognition Competition
source: ILSVRC, ImageNet Large Scale Visual Recognition Competition
The neural networks becomes larger and larger with increased accuracy
32
34. III Proprietary. All rights reserved.
Distributed DNN with Early Exit
Sometimes partial accuracy is
enough – immediate response is
more important
DNN over Distributed Computing
• The DNN is mapped to fog
nodes (device, gateway,
edge server, cloud) that
constitutes the cloud-to-thing
hierarchy
• Each layer in the hierarchy is
responsible for handling parts
of the DNN
• All layers in the hierarchy
work together to make partial
or full inference calculations
source: source: Inverting the Current Cloud-centric Practice: Making End Devices play a Deeper Role in Deep Learning, H. T. Kung, 2016
33
35. III Proprietary. All rights reserved.
DDNN Scaling: Horizontal and Vertical
source: source: Inverting the Current Cloud-centric Practice: Making End Devices play a Deeper Role in Deep Learning, H. T. Kung, 2016
34
36. III Proprietary. All rights reserved.
Not only Micro-cloud
Source: Distributed behavior model orchestration in cognitive IoT solution, Enterprise Information Systems, 2017
Fog as “distributed cloud” -- not only micro cloud, but immersive and
pervasive (over cloud-2-things continuum).
• Most importantly, Fog Should be autonomous and dynamic for the
future IoT Applications and Services.
Combining Distributed DNN with Distributed Clouds (Fogs)
Data-driven behavior models
• Advanced analysis and understanding
• real-time decision that is likely to optimize the business outcome
– dynamically integrate real-world information, conduct what-if
analysis, and choose the course of actions:
(1) get contextual information from the real-world;
(2) need to get deeper analysis (vertically) ?
(3) need to get broader analysis (horizontally) ?
(4) [loop back and again…] perform actuation and get feedback…
35
Source; Talk at the Peking University, "Fog Networking and Computing," T. Russell Hsing, 2017
37. III Proprietary. All rights reserved.
Configurable AI Neural Network(1)
Hierarchical Progressive (Zoom-In) Analysis
• From coarse classification to finer classification
• Example
– Analysis: coarse classification (vehicle)
– Course of actions: collision prevention
– Analysis: fine classification (vehicle models – car, truck, van)
– Course of actions: traffic pattern statistics
– Analysis: More detailed classification (car brands – Toyota, Honda, Ford)
– Course of actions: accident tracking or scene reconstruction
Need to get deeper analysis (vertically) ?
• Drones or robots, not just collect and analyze data for human
• but autonomously evaluate the next actions to take, which may involve
– respond to situations immediately (early exit)
– further AI-inference (maybe in different Fog node) before making decisions
36
Source: III Innovation Project, Frank C. Tsai, 2017
38. III Proprietary. All rights reserved.
Configurable AI Neural Network (1-1)
AI neural network model training/inference in concert (III-Patent pending)
• Model distribution: No need to set up a
complete neural network model, which
consumes chunky memory, for each fine
classification
• Inference: Feature maps of neural
network models may exchange between
fog nodes, per established look-up table,
for further detailed AI inference
• Training: Neural network model for coarse
classification and fine classification model
trained in sequence
– Fine-tuner applies transfer learning based
on coarse classification.
– A look-up table produced indicating from
which layer of coarse classification model,
finer classification is to tap in
Conv-1
ReLU
Conv-2
softmax
FC-2
ReLU
Conv-3
FC-1
ReLU
Conv-1
ReLU
Conv-2
softmax
FC-2
ReLU
Conv-3
FC-1
ReLU
37
Source: III Innovation Project, Frank C. Tsai, 2017
39. III Proprietary. All rights reserved.
Configurable AI Neural Network (2)
38
Horizontal Progressive (Expand-out) Analysis
• From common features to many classifications, sharing the same feature map
• Example
Need to get broader analysis (horizontally) ?
• Face detection for various statistical purpose
• E.g. Gender, age, hair style, glasses-wearing, …
• Multiple “classifiers” may share the same common feature maps, sparing
repeated computations
Various
classifiers
38
Source: III Innovation Project, Frank C. Tsai, 2017
40. III Proprietary. All rights reserved.
Configurable AI Neural Network (2-2)
39
39
Common Feature extraction
Gender classifier
Age classifier
Example: using VGG16,experiments on different layer boundaries
AI neural network model training/inference in concert
• More new classifiers can be independently trained using transfer learning
• Distribute different layers of neural networks to different fog hierarchy (and
with distributed local cache), based on application / business characteristics
Multi-tenancy via virtualization of
oneM2M nodes Male
23
Fog Node layer
boundary (operations
amount )
Fog Node output
data volume
Pool4 (90%) 392 KB
conv5_1 (93%) 392 KB
conv5_2 (96%) 392 KB
conv5_3 (99%) 392 KB
Pool5 (99%) 98 KB
Also, on the same hierarchical level, dynamic distribution of neural network model
layers to different fog nodes per their physical computation resources
39
Source: III Innovation Project, Frank C. Tsai, 2017
41. III Proprietary. All rights reserved.
Edge
Server
Cloud
Gateway
1-1
Gateway
1-2
E W S 北
N
• Nobody shows up
Early exit
• Just passing by
Early exit
• Not target customer
Early exit
• Business
Intelligence
30-min
limited
promo-
tion
Feedback
for greater
customer
under-
standing
…
…
Area view
& instant action
Region view
& instant action
Long-term Trend
& incremental
model training
Virtualization
via containers
Virtualization
via containers
Fog-Fog/Fog-Cloud Orchestration
• Dynamic load balancing
and scaling
Use Case 1 – Smart Retails
Layered
Architecture View
of an IoT System
Example Business Logic
Re-enforcement
Learning
40
Source: III Innovation Project, Frank C. Tsai, 2017
42. III Proprietary. All rights reserved.
Cloud Data Center
+ global compute
+ massive data processing
+ persistent data storage
+ complex decisions
Orchestrator Fog Node
+ local compute
+ data cache
+ temporal data storage
+ simple decisions
Edge-to-Cloud Communication Edge-to-Fog Communication
Fog-to-Cloud Communication
Fog-to-Local Device Communication
Cloud-to-Remote Device Communication
Use Case 2 – AR/VR Shopping
Source: 2017 Asia Fog Summit, III-NCU, Roy T. Wang, 2017/12
41
Shopping Store AR/VR Shopping
TV Shopping Channel On-Line Shopping
43. III Proprietary. All rights reserved.
Use Case 3 - Autonomous Driving
Localization – Key challenge for self-driving
• localization means find the location of car at cm-level precision to make turns
according to navigation, stop and go following the traffic light, changing
lanes, ...
Source: 2017 Asia Fog Summit, ARM, 2017/12
42
44. III Proprietary. All rights reserved.
Embedded Neural Network (1)
43
AI on Edge for embedded or mobile systems
• GPU way too expensive in such systems
• SoC solutions
– Nvidia, Jetson Tegra X1 Super Chip, With 256
Maxwell cores and 4K video capability
– Movidius (acquired by Intel), Fathom Neural
Compute Stick, a modular, deep-learning
accelerator in USB3.0 stick with Myriad 2 VPU
(vision processing unit) for US$79
– Qualcomm, Mediatek, Huawei, AMD and startups
• Google AIY (do-it-yourself artificial intelligence) for makers
• Vision Kit
– Release 2017/12,
ca US$45
– Circuit board with
build-in VPU
– 3 Tensorflow object
recognition models
• Voice Kit
– Release 2017/5
US$60
– Simple version of
GoogleHome
– Built-in Google
Assistant SDK,
Google Cloud Speech
43
45. III Proprietary. All rights reserved.
DNN
Training
Tool
DNN
Optimized
Tool
<<Embedded DL Training Platform>>
(Off-line) Supermicro w/NVIDIA PNY
GeForce GTX 980 4GB
Training
Dataset
developers
Compressed
Models
<< Embedded Real-time Inference system >>
(On-line) nVidia TX1, MediaTek X20
Embedded
Inference Engine
(embedded systems)
(forward , backward training)
Smart devices
Model
Repository
• Images
• Voice
• Sensor Data
• Training Data
• Parameters
Data
Label
Tool
Embedded Neural Network (2)
Software-optimized neural network solutions
• Balance between complexity and accuracy requirements – very applications
dependent
• Algorithm Improving
– FFT based convolution
– Winograd's minimal filtering
– Low Precision GEMM
• Neural Network Architecture Adaptation
– Pruning
– Restructuring
– Quantization
44
46. III Proprietary. All rights reserved.
Model Repository
Model zoo for deep learning image analysis
• Classification – SqueazNet, GoogleNet, ResNet
• Detection – SSD(Single Shot MultiBox Detector), Faster-RCN
• Identification – Re-ID, Openface, Dlib
• Segmentation – SegNet
• ….
• ….
45
47. III Proprietary. All rights reserved.
Distributed Resource Mgmt(1)
46
Overlay logical service layers
• To ease connections between nodes
• To distribute analytic AI-modules
• To dynamically instantiate more or reduce some logical nodes to
reflect computational demand
46
Cloud
Fog
Fog
Fog
Things
Fog
Fog
Things
IN-
CSE
MN-
CSE
MN-
CSE
MN-
CSE
Things
MN-
CSE
MN-
CSE
Things
• Constructing IoT Fog Nodes
based on oneM2M and Docker
– A Fog Node can be mapped to
a Middle Node in oneM2M.
– However, research is required
to extend the oneM2M
architecture to enable multiple
layers of Middle Nodes,
communications among
Middle Nodes and
containerization of Middle
Nodes. Source: 2017 Asia Fog Summit, Joseph Lin , 2017/12
Example – using oneM2M middle nodes to host analytic AI-modules
48. III Proprietary. All rights reserved.
Distributed Resource Mgmt(2)
47
oneM2M -- A Distributed Application Platform for IoT/M2M
47
Source: oneM2M, 2015
Cloud
e.g.
WWAN
oneM2M
messages
e.g. WLAN,
Zigbee
e.g. BLE
Infrastructure
Node
Middle Node
Application
Service
Node
AE
CSE CSE
oneM2M
messages
CSE
AE
API
AE
API
API
Application
Dedicated
Node
AE
API
Cloud
Fog
Fog
Fog
Things
• AE: Behavior model/ decision
• CSE: generic Analytics
• Middle Node to be
dynamically instantiated
49. III Proprietary. All rights reserved.
Load Balancing/Scaling
AI-in-Cloud -- Only a restricted class of
IoT/IIoT applications that are compatible
with the cloud centric model
source: PrismTech (an ADLink co.), 2017/11
AI over Cloud-to-Thing Continuum -- In
general the entities defining an application
distributed across the three tiers
• Virtualizing the infrastructure end-to-end as challenges -- load balancing / scaling
has to be considered end-to-end and should include both Fog and Cloud
resources
– Balancing AI-computation load among Fog nodes via a load balancer that
constantly monitors the load on each Fog node
– Adaptively scaling up and down the deployed Fog nodes, via designing a Fog
manager node that can oversee Fog worker nodes and decide when to scale
up or down the number of Fog workers.
48
50. III Proprietary. All rights reserved.
Concluding Remarks
Fog/cloud hybrid architecture will be the trend
• AI Neural Network training in the Cloud; Inference in the Fog Nodes
• AI Big Data Analytic for long-term trending in the Cloud; Refined inference
model provisioned to the Fog Nodes
Machine Learning (ML) algorithm re-architected for distributed architecture
• Decompose chunky analytic modules into “micro-services” to enable
flexible combination for specific application needs
• Support fast fog-based response and progressively detailed analytics (e.g.
classification) for successive refinement need (in different fog nodes)
• Domain specific algorithms and neural network designs in concert (e.g.
oneM2M decouples AE, CSE) to achieve flexibility and service sharing
• operational management tool and AI modules to design-in
More intelligence moved to the edge
• Increasing number of embedded processing platforms with NN re-architect
• Reducing privacy threat with data being processed locally or proximally
Orchestration among Fog-Fog/Fog-Cloud is still essential (AI tapping in or not)
• Virtualization is to a Fog Node what virtualization is to Cloud
• Orchestration achieved through scaling/load-balancing algorithm and
virtualization
49
52. III Proprietary. All rights reserved.
3 Mega Trend for Gartner
emerging Technologies
Source: Gartner, 2017
51
53. III Proprietary. All rights reserved.
Garner Hype Cycle
52
Source: Gartner, 2017
54. III Proprietary. All rights reserved.
Cognitive Computing
• Cognitive computing (CC) describes technology platforms that, broadly speaking, are based on the scientific
disciplines of artificial intelligence and signal processing. These platforms encompass machine learning,
reasoning, natural language processing, speech recognition and vision (object recognition), human–computer
interaction, dialog and narrative generation, among other technologies.
• At present, there is no widely agreed upon definition for cognitive computing in either academia or industry.
• In general, the term cognitive computing has been used to refer to new hardware and/or software that mimics
the functioning of the human brain (2004) and helps to improve human decision-making. In this sense, CC is a
new type of computing with the goal of more accurate models of how the human brain/mind senses, reasons,
and responds to stimulus. CC applications link data analysis and adaptive page displays (AUI) to adjust
content for a particular type of audience. As such, CC hardware and applications strive to be more affective
and more influential by design.
• Some features that cognitive systems may express are:
• Adaptive: They may learn as information changes, and as goals and requirements evolve. They may resolve
ambiguity and tolerate unpredictability. They may be engineered to feed on dynamic data in real time, or near
real time.
• Interactive: They may interact easily with users so that those users can define their needs comfortably. They
may also interact with other processors, devices, and Cloud services, as well as with people.
• Iterative and stateful: They may aid in defining a problem by asking questions or finding additional source
input if a problem statement is ambiguous or incomplete. They may "remember" previous interactions in a
process and return information that is suitable for the specific application at that point in time.
• Contextual: They may understand, identify, and extract contextual elements such as meaning, syntax, time,
location, appropriate domain, regulations, user’s profile, process, task and goal. They may draw on multiple
sources of information, including both structured and unstructured digital information, as well as sensory inputs
(visual, gestural, auditory, or sensor-provided).
53
55. III Proprietary. All rights reserved.
Cognitive vs Traditional
Applications – the differentiators
Source: Gartner, 2017
54
Reasoning
They reason. They can
understand information but also
the underlying ideas and
concepts. This reasoning ability
can become more advanced over
time. It’s the difference between
the reasoning strategies we used
as children to solve
mathematical problems, and
then the strategies we
developed when we got into
advanced maths like geometry,
algebra and calculus.
Learning
They never stop learning. As a
technology, this means the
system actually gets more
valuable with time. They develop
“expertise”. Think about what it
means to be an expert- - it’s not
about executing a mathematical
model. We don’t consider our
doctors to be experts in their
fields because they answer every
question correctly. We expect
them to be able to reason and be
transparent about their
reasoning, and expose the
rationale for why they came to a
conclusion.
Understanding
Cognitive systems
understand like humans do,
whether that’s through
natural language or the
written word; vocal or
visual.
56. III Proprietary. All rights reserved.
Ambient Computing
• More and more people make decisions based on the effect their actions will have on their own
inner, mental world.[vague] This experience-driven way of acting is a change from the past when
people were primarily concerned about the use value of products and services,[dubious – discuss] and
is the basis for the experience economy. Ambient intelligence addresses this shift in existential
view by emphasizing people and user experience.
• The interest in user experience also grew in importance in the late 1990s because of the
overload of products and services in the information society that were difficult to understand and
hard to use. A strong call emerged to design things from a user's point of view. Ambient
intelligence is influenced by user-centered design where the user is placed in the center of the
design activity and asked to give feedback through specific user evaluations and tests to
improve the design or even co-create the design together with the designer (participatory
design) or with other users (end-user development).
• In order for AmI to become a reality a number of key technologies are required:
• Unobtrusive hardware (miniaturization, nanotechnology, smart devices, sensors etc.)
• Seamless mobile/fixed communication and computing infrastructure (interoperability, wired and
wireless networks, service-oriented architecture, semantic web etc.)
• Dynamic and massively distributed device networks, which are easy to control and program (e.g.
service discovery, auto-configuration, end-user programmable devices and systems etc.)
• Human-centric computer interfaces (intelligent agents, multimodal interaction, context
awareness etc.)
• Dependable and secure systems and devices (self-testing and self repairing software, privacy
ensuring technology etc.
55
57. III Proprietary. All rights reserved.
Digital Twin
• Digital twin refers to a digital replica of physical assets, processes and systems that can be used for various
purposes.[1] The digital representation provides both the elements and the dynamics of how an Internet of
Things device operates and lives throughout its life cycle. [2]
• Digital twins integrate artificial intelligence, machine learning and software analytics with data to create living
digital simulation models that update and change as their physical counterparts change. A digital twin
continuously learns and updates itself from multiple sources to represent its near real-time status, working
condition or position. This learning system, learns from itself, using sensor data that conveys various aspects
of its operating condition; from human experts, such as engineers with deep and relevant industry domain
knowledge; from other similar machines; from other similar fleets of machines; and from the larger systems
and environment in which it may be a part of. A digital twin also integrates historical data from past machine
usage to factor into its digital model.
• In various industrial sectors, twins are being used to optimize the operation and maintenance of physical
assets, systems and manufacturing processes.[3] They are a formative technology for the Industrial Internet of
Things, where physical objects can live and interact with other machines and people virtually.[4]
• An example of how digital twins are used to optimize machines is with the maintenance of power generation
equipment such as power generation turbines, jet engines and locomotives.
• Another example of digital twins is the use of 3D modeling to create digital companions for the physical
objects.[5][6][7] It can be used to view the status of the actual physical object, which provides a way to project
physical objects into the digital world.[8] For example, when sensors collect data from a connected device, the
sensor data can be used to update a "digital twin" copy of the device's state in real time.[9][10][11] The term
"device shadow" is also used for the concept of a digital twin.[12] The digital twin is meant to be an up-to-date
and accurate copy of the physical object's properties and states, including shape, position, gesture, status and
motion.[13]
• A digital twin also can be used for monitoring, diagnostics and prognostics to optimize asset performance and
utilization. In this field, sensory data can be combined with historical data, human expertise and fleet and
simulation learning to improve the outcome of prognostics.[14][15][16][17] Therefore, complex prognostics and
56
58. III Proprietary. All rights reserved.
NIST Fog Definition
Source: Gartner, 2017
57
2017/08/21 - NIST Special Publication 800-191 (Draft).
59. III Proprietary. All rights reserved.
Integrating Fog and MEC
58
MEC Architecture
Fog Architecture
Integrated Architecture
60. III Proprietary. All rights reserved.
IoT, Cloud, and 5G Are Converging
59
IoT
Cloud
5G
• Move computing closer to users
• Make use of distributed resources
• Support multi-tenancy
• Closer interoperability and
integration of applications
• Support diverse devices, apps,
environments, …
• Manage distributed resources
• Secure distributed resources
Common Functionalities
IoT
Cloud
5G
• Virtualization, NFV, Container,
Micro-services, …
• Architecture for moving computing
functions around
• Service automation technology
• Ways to use distributed resources
• Ways to support multi-tenancy
• Lifecycle management of devices,
apps, resources, and systems
• E2E security
Common Technologies
None of these is
unique to Cloud,
IoT, or 5G