Determinants of health, dimensions of health, positive health and spectrum of...
Enabling Smart Manufacturing with Federated AI at the Edge
1. Enabling Smart Manufacturing with Federated AI at
the Edge
Bin Cheng (bin.cheng@neclab.eu)
NEC Labs Europe, Heidelberg, Germany
2. Our Vision: Smart Manufacturing with Edge AI
1
cloud-edge
physical world, lots of moving or fixed edges with cameras and sensors
cloud
edge edge edge
Situation
aware APPs
edge
React on real-time
situations
Customers The knowledge required by applications can be
automatically derived and optimized at cloud and edges
towards the overall business objectives from customersBusiness
objectives
Situation
aware APPs
Situation
aware APPs
Global
knowledge
Regional
knowledge
Local
knowledge
Reacting on
real-time situations
3. Two cases for Edge AI: Distributed AI and Federated AI
2
cloud
e1 e2 e3
Domain A Domain B
cloud
e1 e2 e3
Distributed AI within
single trusted domain
Federated AI across multiple
untrusted domains
5. Fog Function in FogFlow: Serverless Fog Computing
4
Data-centric fog function for dynamic composition of your AI pipelines
6. Adaptive AI Pipelines over Cloud and Edges
5
cloud
edg
e
edg
e
edge
cable
4G
wifi
wifi
# of possible configurations
• exponentially increases as the number of
tasks and their parameters
Deployment must be changed according to the
runtime environment and the current workload, in
order to fulfill the expected QoS
Mauricio Fadel Argerich, Bin Cheng, and Jonathan Fürst. "Reinforcement Learning based Orchestration for Elastic Services." WF-I
Reinforcement Learning based Orchestration
7. Federated AI over Multiple Domains
6
Domain A
Domain
experts
Domain B
Domain experts
Researcher
Domain C
samples
dataset A dataset B
system A system B
classifierclassifier raw data raw data
Function
Catalog
Publish
label functions
Publish/update
label functions
Publish/update
label functions
Select label functions
and also exchange
relevant models
Label
functions
Select label functions
and exchange
relevant models
8. Function Catalog: The Zoo of Data Processing Functions
7
data Sampling cleaning annotating labelling training predicting decisions
Data preparation (80% of the total effort)
9. Data Usage Control cross Domains
8
HyperLedge/Blockchain
(accountability)
data AI
functions
Usage
policy
Usage
evidence
data
data
AI
functions
AI
functions
Honest but curious
(semi-trusted model)
follows protocol but will try to learn
as much information as possible,
without actively “cheating”
10. Use Case: Defect Product Detection for Smart Factory
9
Workbench NWorkbench A
Marketplace
for AI Models
Defect Product
Detection
Local AI model creation
(labelling, training,
and learning at edges)
Sharing of AI models
for defect goods
Local AI model
federation and
usage
……