embedding artificial intelligence into our lives · recognition and translation, vision, and more...
TRANSCRIPT
© 2018 Synopsys, Inc. 1
Embedding Artificial Intelligence into Our Lives
Michael Thompson, SynopsysD&R IP-SOC DAYS Santa Clara
April 2018
© 2018 Synopsys, Inc. 2
Agenda
Introduction
What AI is and is Not
Where AI is being used
Rapid Advance of AI
Summary
D&R IP SOC DAYS - April 2018
© 2018 Synopsys, Inc. 3
Artificial Intelligence
Moving from Mainframe to Embedded
Evolving Rapidly
Requires very high-performance coupled with low power and cost
Today encompasses range of applications from search in the cloud, to cars, robotics, games, speech recognition and translation, vision, and more
Artificial intelligence (AI) is the enabling of a machine to perceive its environment and respond in a way that increases its usefulness to us.
It’s Been Around for 60 years
John McCarthy coined the term in 1956 at a Dartmouth conference
D&R IP SOC DAYS - April 2018
© 2018 Synopsys, Inc. 4
Realizing AIs Full PotentialHumans are Fallible Drivers
Annual Global Road Crash Statistics• 1.3 million people die in road crashes, 3,287 deaths a day.
• Additional 20-50 million are injured
• Road crashes cost USD $518 billion globally
94% of accidents caused by human error– 2% environment, 2% mechanical, 2% margin error
ADAS Systems• Lane keeping, lane change, lane departure
• Pedestrian Detection
• AEB/Automatic Emergency Braking), forward collision
• Adaptive cruise control
• Traction and stability control
• Blind spot monitor
It will take years to realize the full potential of AIIt will take years to realize the full potential of AI
7%
2%
5%
L4
L2L3
L1L0
L5
2005 2015 2025 2035
0
5
10
15
20
25
Mill
ion
s
Global Autonomous Vehicle Sales Forecast (L4+L5)
D&R IP SOC DAYS - April 2018
© 2018 Synopsys, Inc. 5
What Does This Look Like in 20 Years?We Are Entering the AI Era
7G
6G
5G
AI will increase productivity, data access, safety and change how we interact, work, liveAI will increase productivity, data access, safety and change how we interact, work, live
D&R IP SOC DAYS - April 2018
© 2018 Synopsys, Inc. 6
Artificial Intelligence and Deep Learning
• Artificial Intelligence
– Narrow AI (weak AI): Technology outperforming humans in a narrowly defined task
– Artificial General Intelligence (strong AI): Human levels of intelligence exhibited by machines
• Machine Learning
– Application of AI uses algorithms to analyze data and infers information about real world
• Neural Networks
– Class of machine learning algorithms modeled after the human brain
– Neuron represents the computational unit, network describes how units are connected
• Deep Learning / Deep Neural Networks
– A subset of machine learning using artificial neural networks
– Deep neural networks are capable of learning using large data sets
D&R IP SOC DAYS - April 2018
© 2018 Synopsys, Inc. 7
From Computer Vision
• Large scale server farms• High cost• High energy use• Large footprint
• On-chip• Cost effective• Energy efficient• In your pocket
To Embedded Vision
Computer Vision
Machine Learning
Embedded Systems
AI Moving from Cloud to Embedded
D&R IP SOC DAYS - April 2018
© 2018 Synopsys, Inc. 8
Is AI Intelligence?
• Not everyone considers AI to be intelligence – Sophisticated manipulation of data and our emotions
– This point of view is not unreasonable
• Intelligence can be defined as:
– Ability to perceive the environment and take actions to maximize the chance of success What AI is today
– Can also be defined as the skilled use of reason What humans do
AI will eventually encompass the ability to reason and will likely eclipse human intelligence
But ability to reason and human intelligence are very complex processes
D&R IP SOC DAYS - April 2018
© 2018 Synopsys, Inc. 9
AI Today
• When we think of AI we tend to think of humanoid machines
• Reality is more Amazon Echo– Combines voice recognition (perception), fast processing (decision making),
and an action (response)
• Perception: Sensors, cameras, a database, spoken request or other sources
• Processing: local processor, in the cloud, or both to increase performance
• Response: audio, mechanical, database update, visual, something else
Good AIBad AI
D&R IP SOC DAYS - April 2018
© 2018 Synopsys, Inc. 10
Levels of AI in Use Today
• Low-end - Chess games – Use brute force to analyze all moves with next move based on a series of moves with
the highest chance of winning (Deep Blue, DeepChess)
• Mid-range – Object recognition and classification– Requires an understanding (training) of what is being looked for
• High-end - Language translation
– Requires an understanding of word structure and context in the language that the words are being translated to and from
– More than brute force computation
• These are all Narrow AI (weak AI) and do not require human levels of thinking
D&R IP SOC DAYS - April 2018
© 2018 Synopsys, Inc. 11
Definition of What is AI Changes Over Time
• Tasks that were once defined as AI have been removed from the list– Optical character recognition and expert systems
– No longer considered AI because they are considered routine
• A List of things that are generally considered to be AI in 2018– Competing at a high level in a strategic game (chess and Go)
– Understanding language
– Interpreting complex data
– Intelligent routing in content delivery networks
– Autonomous vehicles
– Machine vision
• List will change over time – Due to advancements in AI
– As applications become routine
D&R IP SOC DAYS - April 2018
© 2018 Synopsys, Inc. 12
Machine Vision
• Has been around for years
• Evolving and moving to embedded
• Machines now achieve higher levels of accuracy than human experts
• Orders of magnitude faster than humans
• New algorithms are faster and more accurate
• Scene segmentation can be done on HD video at 60 fps
• The results are truly amazing
D&R IP SOC DAYS - April 2018
© 2018 Synopsys, Inc. 13
YOLO v2 Video
Skyfall YOLO v2 Video
D&R IP SOC DAYS - April 2018
© 2018 Synopsys, Inc. 14
Neural Networks
• Widely used for machine vision
• Have dramatically increased accuracy– Mimic the way our brain learns
– Uses information and training to recognize patterns
– New algorithms are faster, more accurate, and simpler
• Used in other applications too– Character recognition
– Text generation
– Language translation
– Audio
• NASA uses NN to analyze data from telescopes– More accurate than humans and much faster
– Recently found an 8th planet revolving around Kepler-90 that is 2545 lights years away
– First known solar system with 8 planets outside of our own
D&R IP SOC DAYS - April 2018
© 2018 Synopsys, Inc. 15
Implementing AI in Embedded Applications• Being facilitated by advancements in microprocessor capabilities
• Combined with advancements in process technology
• Enabling very small processors with performance levels that were unattainable a few years
• ARC HS cores deliver up to 7500 DMIPS per core, fit into 0.06mm2 and use 50uW/MHz power
• Can be scaled to even higher performance with dual-core and quad-core versions
AI development platform using ARC HS that can be used for various AI applications developed by NARL in Taiwan http://www.cic.org.tw/aisoc/aisoc.jsp
D&R IP SOC DAYS - April 2018
© 2018 Synopsys, Inc. 16
Specialized Embedded Vision Processors
• Offer the highest performance for embedded applications
• Target vision applications and use Neural Network capabilities
• Replacing GPUs in many applications
• Synopsys EV6x processors deliver up to 4.5 Tera MACs per second
• Support for full range of CNN algorithms– AlexNet, GoogLeNet, ResNet, SqueezeNet,
TinyYolo, Yolo v2 and others
• High productivity standards-based toolset– OpenCV libraries, OpenVX framework,
OpenCL C compiler, C/C++ compiler and CNN mapping tools
Vision CPU (1/2/4 cores) CNN Engine (scalable)
AXI Interconnect
Synopsys EV6x Embedded Vision Processor
Shared MemoryShared MemorySync & DebugSync & Debug Streaming Transfer UnitStreaming Transfer Unit
Libraries (OpenCV) &
API (OpenVX)
Libraries (OpenCV) &
API (OpenVX)
Simulators (fast NSIM,
EV VDK)
Simulators (fast NSIM,
EV VDK)
Compilers / Debuggers (C/C++,
OpenCL C)
Compilers / Debuggers (C/C++,
OpenCL C)
MetaWare EV
CNN Mapping
Tool
CNN Mapping
Tool
Core 4Core 4
Core 3Core 3
Core 2Core 2
Core 1Core 1
32-bit scalar
512-bit vector DSP
VFPU
3520 MAC Engine3520 MAC Engine
1760 MAC Engine1760 MAC Engine
880 MAC Engine880 MAC Engine
ConvolutionConv. 2D
ConvolutionConv. 2D
ClassificationConv. 1D
ClassificationConv. 1D
SFPU
D&R IP SOC DAYS - April 2018
© 2018 Synopsys, Inc. 17
Dramatic Algorithm ImprovementObject Classification with CNNs
VGG16 GoogleNetv1AlexNet
2012
Classification (1000)
8 layers, 15.4% errorComing out party for deep
learning techniques at ILSVRC
2014
Classification (1000)
16-19 layers, 7.3% errorSimple 3x3 convolutions and
deeper layers
CNN Graphs for ClassificationCNN Graphs for Classification
2014
Classification (1000)
22 layers, 6.7% errorIntroduced the idea that CNN layers didn’t always have to be stacked up
sequentially
Resnet
2015
Classification (1000)
152 layers, 3.6% error
D&R IP SOC DAYS - April 2018
© 2018 Synopsys, Inc. 18
The AI Era Begins
• Interesting to see how AI develops over the next 10 years– Cars will drive themselves,
– Personal assistants will be a great deal more clever
– Seamless natural language translation
• Amazing new AI applications that haven’t been thought of yet
• Enabled by advanced processors like Synopsys EV6x family
• We are on the leading edge of the era of artificial intelligence and just starting to see the capabilities
• AI won’t replace us as some fear, but as it evolves over the coming years it will have a profound impact on our lives
Vision CPU (1 to 4 cores)Vision CPU (1 to 4 cores)
AXI InterconnectAXI Interconnect
Core 4Core 4
Core 3Core 3
CNN EngineUp to 3520 MACs
CNN EngineUp to 3520 MACs
Core 2Core 2
Core 1 Core 1
32-bit scalar
512-bit vector DSP
Synopsys EV6x Embedded Vision Processor
ConvolutionConvolution
ClassificationClassification
32-bit scalar
512-bit vector DSP
Shared MemoryShared MemorySync & DebugSync & Debug Streaming Transfer UnitStreaming Transfer Unit
D&R IP SOC DAYS - April 2018
Thank You