Making Sense of Artificial Intelligence:A Practical Guide
JEDEC Mobile
& IOT Forum
Copyright © 2018 Young Paik,Samsung Senior Director Product Planning
Disclaimer
• This presentation and/or accompanying oral statements by Samsung representatives collectively, the “Presentation”) is intended to provide information concerning the SSD and memory industry and Samsung Electronics Co., Ltd. and certain affiliates (collectively, “Samsung”). While Samsung strives to provide information that is accurate and up-to-date, this Presentation may nonetheless contain inaccuracies or omissions. As a consequence, Samsung does not in any way guarantee the accuracy or completeness of the information provided in this Presentation.
• This Presentation may include forward-looking statements, including, but not limited to, statements about any matter that is not a historical fact; statements regarding Samsung’s intentions, beliefs or current expectations concerning, among other things, market prospects, technological developments, growth, strategies, and the industry in which Samsung operates; and statements regarding products or features that are still in development. By their nature, forward-looking statements involve risks and uncertainties, because they relate to events and depend on circumstances that may or may not occur in the future. Samsung cautions you that forward looking statements are not guarantees of future performance and that the actual developments of Samsung, the market, or industry in which Samsung operates may differ materially from those made or suggested by the forward-looking statements in this Presentation. In addition, even if such forward-looking statements are shown to be accurate, those developments may not be indicative of developments in future periods.
2
Deep Learning Has Changed AI
Artificial
Intelligence
Machine
Learning
Deep Learning
Data DrivenRules Based
Not accurate enough to be
commercially viableApproaching Human
Level Accuracy
Example: Speech recognitionFew words ~ 99% accurate
Many words ~ 60% accurate
This Is Not That Kind of Talk
Stochastic Gradient Descent
Backpropagation
Variable precision math
Deep Neural Net
Convolutional Neural NetRecurrent Neural Net
LSTM
Softmax
Pooling
ReLu
Supervised Training
Deep Learning has a lot of jargon…
This talk will skip as much as possible.
Fully connected
GPU
CUDA
OpenCL
High-Level AI Flow
Data
Storage
Data
Processing
DL Model
Deployment
SSD
DDR
GDDR
HBM
Mobile CPU
NPU
LPDDRGPU Memory
Many TB - PB+
>1 TBps BW
Deep Learning Simplified
< 1 MB – 10 GB
Phones AIAssistants
HomeDevices IoT
Auto
Training Inference
Hours - Weeks
3 Takeaways
�Data is King
More data = more accuracy
�Deep Learning is hard
Leave it to the professionals
�You don’t have to use one AI
Many, smaller AIs are better than one big one
Circle of (DL) Life
DL models need to be constantly be fed data.
When designing new products find ways to feed back data to improve AIs.
Gather Data
Create AI
DeployExample:
Home thermostat:User override of settings should be
seen as an AI fault and feed back.
Real World Considerations
Data Training DL Model
Network Device
Server
Training DurationTraining may take days or even weeks. Hard to parallelize.
DL Model SizeIn general, larger models are more accurate, but may be hard to fit on mobile/IoT devices.
LatencyKeeping models on servers will increase latency.
PrivacyConstantly feeding data back to servers may have privacy concerns.
PowerConstantly running DNNs on devices may take a lot of power.
BandwidthEnd users may not want to constantly use bandwidth to servers.
ComputeMobile/IoT resources may not be enough to run on-device.
DL Model on-device
DL Model on-server
Decentralized DBSome apps require additional data that are best centrally stored.
Q: How would you implement mobile speech recognition?
Tricks For Improvement
Transfer Learning Model Compressions
Take a pre-trained DL model and retrain it with new data.
Pros• Trains much faster
than first model.
• No need for AI
scientist.
• Easy frameworks
available in cloud.
Cons• Use case must be
very similar.
• Not as accurate as
original.
Pros• DL models may
compress by 500x.
• Accuracy may not be
impacted.
• May decrease power
needed to infer.
Cons• Requires additional
processing after
initial training.
Many weights and nodes may not be important and can be removed.
Real World Considerations
Data Training DL Model
Network Device
Server
Training DurationTraining may take days or even weeks. Hard to parallelize.
DL Model SizeIn general, larger models are more accurate, but may be hard to fit on mobile/IoT devices.
LatencyKeeping models on servers will increase latency.
PrivacyConstantly feeding data back to servers may have privacy concerns.
PowerConstantly running DNNs on devices may take a lot of power.
BandwidthEnd users may not want to constantly use bandwidth to servers.
ComputeMobile/IoT resources may not be enough to run on-device.
DL Model on-device
DL Model on-server
Decentralized DBSome apps require additional data that are best centrally stored.
Q: How would you implement mobile speech recognition?
Find New Ways Of Using Old Tech
Artificial
Intelligence
Machine
Learning
Deep Learning
Data DrivenRules Based
New uses?Approaching Human
Level Accuracy
Example: Speech recognitionFew words ~ 99% accurateMany words ~ 60% accurate
Making Use of Multiple AIs
Data Training Model
Server
Network Device
PowerSpecialized AI for hotwordsrequire much less power.
Privacy and BandwidthUse of hotword minimizes data sent back to server.
LatencyLatency for full AI is still longer, but now hotword recognition masks issue.
Decentralized DBCentralized DB can store data that can be used by AI.
“HOTWORD, What is …”“What is …”
Other Use Cases
Data Training DL Model
Network Device
Server
DL Model on-device
DL Model on-server
Q: How would you implement …
Autonomous Driving?
Security Cameras?Facial
Recognition?
Drone Delivery
Conclusions
• AI is still early in its development.
• Design of AI systems is evolving.
• You may find new uses for old ideas.