Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Overview
Background
Intelligence at the Edge
Samsung Research
Learning at the Edge: Challenges and Brainstorming
Amazon Alexa Smart Home!
Background
Ph.D. at UW CSE
RFID, Mobile, Sensors, Data
Nokia Research Samsung Research Silicon Valley
Context Framework
On-Device Analytics
AlgoSnap
Machine Learning at Amazon Alexa Smart Home
Intelligence at the Edge
“Edge”
End-points that generate data
Social, Financial, Physical Sensors,
Environmental Sensors, Health Sensors,
Car, Home, Television, Media Consumption,
Searches, Location, Calendars, Purchases
. . .
All Data
All Data
All Intelligence
Shared back
to devices
. . .
All Data
All Data
All Intelligence
Shared back
to devices
Why Not?Energy, latency, privacy…
IoT Increased Volume, Variety, Velocity!
Radio
Energy
Network
Latency
Privacy
Smartphone Data Type Sensor Types Estimated Avg
Location GPS, WiFi, Cell Towers, Bluetooth 40K / min
Device Motion Accelerometer, Gyroscope, Compass 160K / min
Environmental Camera, microphone, Light, proximity, temperature, pressure, magnetic field
37MB / minno video:
(960K / min)
Device Interaction keys pressed, touch screen, App usage, media usage, screen on/off, etc
20K / min
Social Calls, SMS, emails, Facebook, Twitter, Calendar, contacts
1K / min
Interest/content Browser, search, purchases, bookmarks 20K / min
Wearables Sensor + interaction data from wearables 160K-180K / min
Purchases Web transactions, NFC transactions 1K / min
1.5 to 40 MB/min per user ~1 Exabyte/day at Facebook Scale
Intelligence at the Edge
Save Energy: Push computation to the data
Reduce Latency: Run models on the user’s device
Enhance Privacy: Don’t upload data
Example: Centaurus - Edge Framework
Concept: shift data and processing to the device-side
. . .
User
1 User
2User
3
User
N
User
N+1
User
N+2
. . .User
1 User
2User
3
User
N
User
N+1
User
N+2
Example: Centaurus - Edge Framework
Concept: shift data and processing to the device-side
. . .
User
1
User
1
User
1
User
1
User
1
User
1
Only High-Level Context is Sent to the Cloud, with User Consent
Example: Centaurus - Edge Framework
Expressive scripts specified as dataflows
Operators transform raw data
Models trained in cloud with big data set
SumSimilarity
Euclidean Dist.
SegmentBand-PassAvg Entropy
Tokenize
CountMax
N-gramMin
Tokenize
MedianNaïveBayes
MagnitudePatternMatch
DifferenceEnergy
Std Dev
DecisionTree
CorrelationFFT
Duration
…
On Device: Intelligence Script Engine
Modules to connect to data sources (sensors, logs, social networks)
Library of data Processors
Example:
Gyro Data FFT of Gyro
FFT
Context Scripts configure Operators
Examples:
Script for “Watching a Movie”
Script for “Walking”
Example: Centaurus - Edge Framework Quantifying savings: Walking Detection on smartphone
20 hours accelerometer @ 10Hz
Implement with Centaurus - uploads only classification
Implement in Cloud – uploads all data
Centaurus uploads only 0.14% of the data
Centaurus time-to-classification is slightly faster
Power consumption (network is biggest power hog):
WLAN Upload Data Size (Kb) Power (mJ) Time (ms) mJ/Kb
Centaurus 27 148 237 0.97
Cloud-Only 55,487 66,827 84,788 0.42
HSPA Upload Data Size (Kb) Power (mJ) Time (ms) mJ/Kb
Centaurus 27 3069 2,896 64
Cloud-Only 55,487 1,048,766 545,826 15
The Next Step: Learning at the Edge
Federated Learning – Google Research Blog – April 6, 2017https://research.googleblog.com/2017/04/federated-learning-collaborative.html
Push training to the Edge, not just models
Raw data never leaves the Edge!
Challenges and brainstorming:
1) Decentralized learning: general + personal models
The Next Step: Learning at the Edge
Push training to the Edge, not just models
Raw data never leaves the Edge!
Challenges and brainstorming:
1) Decentralized learning: general + personal models
2) Adapting to varying device resources
3) Security and privacy between device and cloud
4) Supervised learning: soliciting user labels
The Next Step: Learning at the Edge
Push training to the Edge, not just models
Raw data never leaves the Edge!
Challenges and brainstorming:
1) Decentralized learning: general + personal models
2) Adapting to varying device resources
3) Security and privacy between device and cloud
4) Supervised learning: soliciting user labels
5) Peer-to-Peer coordination at the Edge
Amazon Alexa Smart Home
200 Engineers and Scientists and growing fast
Petabyte-scale data: millions of customers & devices
Hiring Engineers and Scientists at all levels!
Contact: [email protected]
Related Documents
