For the full video of this presentation, please visit:
https://www.embedded-vision.com/platinum-members/embedded-vision-alliance/embedded-vision-training/videos/pages/feb-2017-member-meeting-rowen
For more information about embedded vision, please visit:
http://www.embedded-vision.com
Chris Rowen of Cognite Ventures delivers the presentation "The Vision AI Start-ups That Matter Most" at the February 2017 Embedded Vision Alliance Member Meeting. Rowen shares his unique perspective on the vision AI start-ups that matter most.
TeamStation AI System Report LATAM IT Salaries 2024
"The Vision AI Start-ups That Matter Most," a Presentation from Cognite Ventures
1. The Vision AI Start-ups That Matter Most
Chris Rowen, PhD, FIEEE
CEO
Cognite Ventures
February 22, 2017
2.
3. Outline
• The Vision AI Innovation Scene
• Quick Picture of Cognitive Computing Startups
• Embedded Vision and Cognitive Computing
• Embedded Vision Innovation Scene
• New Hardware for Embedded Vision
• Observations
• The List
4. Vision AI Innovation Scene
• Huge enthusiasm for cognitive computing aka artificial intelligence aka machine
learning aka deep learning aka neural networks
• Transformation of all modeling and recognition of complex data, scenes and actions
• Applications from robotics to human-machine interface
• Profound changes in how we develop systems: programming labeling and training
• Massive investment by major cloud operators: Google, Baidu, Facebook, Microsoft,
Amazon sucking up every deep learning PhD they can find
• Massive startup activity around the world
• Which startups are the most important and focused on serious cognitive computing?
From raw catalog of ~2200 AI startups, we’ve identified ~275 of particular interest – 75
of them in embedded vision
5. Typical omission:
Exceed.ai’s chatbot platform
helps brands increase revenue
and customer loyalty by enabling
personalized one-on-one
conversations with their
customers at scale through any
messaging application.
Typical inclusion:
drive.ai creates AI software for
autonomous vehicles. It aims to
build a hardware and software
kit powered by artificial
intelligence for carmakers.
What applications are at the heart of AI?
Autonomous
Driving
Cyber
Security
Surveillance
Generic
deep
learning
platforms
Medical
imaging
Massive
Scale
Sensor IoT –
e.g. Ag
Legal
Analysis
Advertising
Automation
Drones
Toys
Robotics
Generic
Data
Analytics
CRM
Language
Translation
Financial
modeling
Predictive
Marketing
Geoanalytics
ERP
Retail
AnalyticsSelection Criteria:
• Inherent huge data
• Embrace and tight
focus on deep
learning methods
for modeling and
classification
• Job listings and
blogs show depth in
machine learning
6. Quick Picture of Cognitive Computing Startups: What?
Cognitive Computing: 277
Embedded: 82
Vision: 125
Embedded Vision:74
• Almost ¾ of 276 startups focus on cloud
software: CRM, logistics, predictive
marketing
• Heavy emphasis on document and text
processing in the cloud
• Plenty of vision/image processing in the
cloud: non-real-time analysis
• Most of embedded includes vision. Rest
is audio, voice and motion sensing
7. Quick Picture of Cognitive Computing Startups: Where?
CA, 67%
MA, 10%
NY, 5%
TX, 3%
WA, 2%
CA MA
NY TX
WA PA
MI CT
DC NM
TN MD
FL OR
UT OH
CO DE
USA, 49%
UK, 21%
China, 8%
Canada, 5%
Germany, 3%
India, 2%
USA
UK
China
Canada
Germany
India
Japan
Israel
The Netherlands
South Korea
France
Switzerland
Spain
Russia
Austria
South Africa
Slovenia
Norway
Ireland
Hungary
Finland
Denmark
Belgium
Argentina
8. Embedded Neural Network Product Segments
Autonomous
Vehicles and
Robotics
Monitoring,
Inspection and
Surveillance
Human-Machine
Interface
Personal Device
Enhancement
Vision Multi-sensor: image, depth, speed
Environmental assessment
Full surround views
Attention monitoring
Command interface
Multi-mode ASR
Social photography
Augmented Reality
Audio Ultrasonic sensing
Acoustic surveillance
Health and
performance
monitoring
Mood analysis
Command interface
ASR social media
Hands-free UI
Audio geolocation
Natural
Language
Access control
Sentiment analysis
Mood analysis
Command interface
Real-time translation
Local bots
Enhanced search
9. Why Cognitive Vision Now?
• Computing and communication driven by new data
in/out
• CMOS sensors trigger imaging explosion
• 99% of of captured raw data is pixels (dwarfs sounds
and motion)
1010 sensors x 108 pixels/sec = 1018 raw pixels/sec
• Rapid growth of vision-based products and services
• Starting 2015: more image sensors than people
• New Age: Making sense of pixels requires computer
cognition 0
2E+09
4E+09
6E+09
8E+09
1E+10
1.2E+10
1.4E+10
1.6E+10
1.8E+10
2E+10
1990 1995 2000 2005 2010 2015 2020
World Population
Three-year sensor
population
10. Vision
0
10
20
30
40
50
2011 2012 2013 2014 2015 2016
ImageNetTop-1Error
%
Year
Rapid Progress on Accuracy
0
10
20
30
40
50
0 5 10 15 20 25
ImageNetTop-1
Error%
GMACs per image
Bounded Compute Load
0
10
20
30
40
50
0 50,000,000 100,000,000 150,000,000
ImageNetTop-1Error
%
Model Coefficients
Models Getting More Manageable
• Computer vision is big, obvious NN domain
• Many related tasks: classification, localization,
segmentation, object recognition, captioning,
generation
• Huge computation in embedded inference
• Vision is fundamentally hard – even for humans!!
• Example: ImageNet Classification:
• 1000 categories
• 120 species of dogs
23.5
24
24.5
25
25.5
0 2 4 6 8
ImageNetTop-1Error
%
GMACs
Optimization Doubles Efficiency
ResNet 50,101
Cadence
Tibetan mastiff Shih-Tzu Norwegian elkhound
11. Embedded Vision Startup Scene: What and Where?
• Identified 74 startups focused on machine learning for embedded vision
• Half in US, >30% in CA. Half doing robots, drones and cars
0
5
10
15
20
25
30
35
Cognitive EV Startup Countries
CO
MI
PA
TX
MA
Surveillance
Vehicles
Human-
Machine
Interface
Drones and
Robots
Silicon
0
5
10
15
20
Application for Cognitive EV Startups
12. Some Examples
DeepGlint (格灵深瞳) [China]
•18M in 3 Rounds, most recent June, 2014
•Deep Glint focuses 3D computer vision and machine learning technologies, to provide
automatic human trajectory analysis solutions to banks and shopping centers.
Mashgin [US]
•$620k in 3 Rounds, most recent September, 2015
•Mashgin is building a self-checkout kiosk that uses 3D reconstruction, computer vision
and deep learning to identify items.
FiveAI [UK]
$2.7M Seed on July, 2016
We're building the world's most reliable autonomous vehicle software stack
to solve the most difficult problem of all - delivering a solution that's safe in
complex urban environments, without any driver involvement.
13. Embedded Vision Drives Hardware
Need scale and efficiency:
• Conventional wisdom - deep neural
networks much less efficient than
hand-tuned feature recognition
methods (but more effective)
• Convolutional neural networks allow
• High parallelism
• Low bit resolution
• Structured, specialized architectures
• Manageable memory bandwidth
• ~1000x energy improvement over GP
CPU may compensate for efficiency gap
10
100
1,000
10,000
10 100 1,000 10,000 100,000
GMACSPerWatt
GMACs
Neural Network Platforms
Vision DSP core 1 Vision DSP core 1 cluster Vision DSP core 2
Vision DSP core 2 cluster Embedded GPU core cluster Data Center GPU 1 cluster
Embedded GPU cluster FPGA 1 FPGA 2
Convolutional neural network (CNN) engine CNN engine cluster Data Center GPU 2 cluster
Data Center GPUs
FPGAs
Embedded
GPUs
Vision DSPs
Vision + NN DSPs
CNN engines
GPCPU
14. Hardware Startups for Embedded Vision
• Modest number of startups reflects challenging funding for silicon
• Rapid pace of change in basic neural network algorithms dictates programmability
• Major semis and IP providers investing heavily in vision and deep learning
Name Description Website Country State
BrainChip Spiking Neuron Adaptive Processor www.brainchipinc.com USA CA
Cambricon Device and cloud processors for AI www.cambricon.com China
Cerebras Systems Specialized next-generation chip for deep-learning applications cerebras.net USA CA
Deep Vision Low-power silicon architecture for computer vision www.deepvision.io USA CA
Deephi Compressed CNN networks and processors www.deephi.com China
Graphcore Graph-oriented processors for deep learning www.graphcore.ai UK
Isocline Ultra-low power NN inference IC design based on flash+analog+digital www.isosemi.com USA TX
KNUPATH Ultra-scale processor ICs for vision and ML www.knupath.com USA TX
Leapmind Embedded deep learning platform www.leapmind.io Japan
Reduced Energy
Microsystems
Lowest power silicon for deep learning and machine vision www.remicro.com USA CA
Tenstorrent
Deep learning processor: designed for faster training and adaptability to
future algorithms
www.tenstorrent.com Canada
ThinCI vision processing chips www.thinci.com USA CA
15. The Cognitive Embedded Vision List
Abundant Robotics
Accelerated
Dynamics
AIMotive
Airware
AKA
Alchera Technologies
Algocian
Anki
Argo AI
Auro Robotics
Blue Vision Labs
BrainChip
Cambricon
Cerebras Systems
Clearpath Robotics
CloudMinds
Cognitive Pilot
Comma AI
Deep Vision
Deep Vision
DeepGlint
Deephi
DeepScale
Drive.ai
Emotibot
Emovu
Emteq
Evolve Dynamics
Face++
FiveAI
Graphcore
Horizon Robotics
Intuition Robotics
Iris Automation
Isocline
Kindred
Kneron
KNUPATH
Leapmind
Lily Camera
Machines with Vision
Mashgin
Memkite
Minieye
Momenta
MorpX
Nauto
Netradyne
Neurala
Novumind
Noxton Analytics
nuTonomy
Osaro
Oxbotica
Pilot AI Labs
Quanergy
Reduced Energy Microsystems
RobArt
RoboCV
Rokid
Scortex
Shield AI
Skydio
Sportcaster
Tenstorrent
TeraDeep
ThinCI
Third Eye Systems
Universal Robotics
Velodyne
Viz
White Matter
Zero Zero Robotics
Zoox
16. Observations
1.Yes, the cognitive startup space is extremely lively, but demand exceeds
supply – many under-served niches
2.Deep learning disrupts embedded vision – massive retraining of people and
retooling of solutions
3.New technology enables new business models, especially
• new services – e.g. “emotion as a service”
• cloud-device hybrids – e.g. base layers trained in cloud, transfer learning in device
• human augmentation – e.g. guided labeling tools, remote supervision data programming
• autonomous agents – e.g. knowledge structures in support of task reasoning (AIBrain)
4.Expertise and grit gets funded
17. Sources and Caveats
• Crunchbase: Comprehensive database of startups but needs lots of sifting: includes unfunded
dreams, research programs, acquired companies, etc.
• Shivon Zilis and James Cham in the Harvard Business Review: “The State of Machine Intelligence,
2016”
• MMC Ventures: “Artificial Intelligence in the UK: Landscape and learnings from 226 startups”
• 机器之心选出全球最值得关注的100家人工智能公司: (“The heart of the machine selected the
world's most noteworthy 100 artificial intelligence companies”)
Caveats:
• Somewhat subjective selection based on estimated degree on concentration on machine learning and
neural network methods
• The scene is constantly changing, with companies appearing and disappearing daily
• Hard to get every link and summary correct
• Send me corrections and suggestions: rowen@cogniteventures.com
18. The Whole Cognitive Computing List
www.cogniteventures.com/the-cognitive-computing-startup-list
4Paradigm
ABEJA
Abundant Robotics
Accelerated Dynamics
Affectiva
AIBrain
AiDO
AIMotive
Airware
AKA
Alchera Technologies
Algocian
Algorithmic Intuition
Alpha I
Amplero
Anki
Argo AI
Arimo
Arimo
Arya.ai
Atomwise
Auro Robotics
Aurora AI
Automat
Avalon AI
Aylien
Bay Labs
Behavox
Behold.ai
Benevolent
BenevolentAI
Birds.ai
Bloomsbury AI
Blue Vision Labs
BMLL Technologies
BrainChip
Butterfly Networks
Calipsa
Camio
Capio
Captricity
CareSkore
Cartoaware
Celaton
Cerebellum Capital
Cerebras Systems
Cirrascale
Citrine Informatics
Clarifai
Clear Metal
Clearpath Robotics
CloudMedx
CloudMinds
Cogitai
Cognative
Cognicor
Cognitive Pilot
Comma AI
Content Technologies
Cortexica
Cortica
Cortical.io
CrossingMinds
CrowdAI
CrowdFlower
Cyberlytic
CyCorp
Cylance
Cyra
DarkTrace
Datalogue
Dataminr
DataRobot
Deep Genomics
Deep Instinct
Deep Vision
Deep Vision
Deep Vision
Deep6 Analytics
DeepGlint
Deephi
Deepomatic
DeepScale*
DeepSense.io
DeepVu
Descartes Labs
Digital Reasoning
DigitalGenius
Ditto Labs
Drive.ai
DroneDeploy
Eloquent Labs
Emotech Ltd
Emotibot
Emotient
Emovu
Emteq
Enlitic
Evolve Dynamics
Eyeris
Face++
Fashwell
FeatureSpace
FiveAI
FoodVisor
Grail Inc.
Graphcore
Graphistry
Greedy Intelligence
GridSpace
H2O
Hocrox
Horizon Robotics
Hubino
Hyperverge
ICarbonX
Idio
Imagia
Imubit
Indico
Insilico Medicine
Intelligent Voice
Intelnics
Intuition Robotics
Iris Automation
iSentium
Isocline
JukeDeck
Kaggle
Keen Research
Kheiron Medical
Kimera
Kindred
Kneron
KNUPATH
KONUX
Last Mile Technologies
LastMile
Leapmind
Level 6 AI
Leverton
Lexalytics
Lily Camera
Linguamatics
Loop AI
Luminist
Luminoso
Lunit
Maana
Machines with Vision
Mad Street Dan
Marax AI
Mashgin
Matroid
MEDANN
MedWhat
Memkite
Mentat Innovations
micropsi
MindMeld
Minds.ai*
MINDSET
Minieye
Mobvoi
Momenta
MorpX
Nara Logics
Nauto
Neo AI
Netra
Netradyne
Neural Painting
Neurala
Neurence
NNaisense
Novumind
Noxton Analytics
Nudgr
Numenta
Numerate
nuTonomy
Oncora Medical
OpenAI
OpenCapacity
Orbital Insight
Osaro
Oseven
Oxbotica
Pat
Phrasee
Pilot AI Labs
PitStop
Pixoneye
Planet
Pop Up Archive
Preferred Networks
Prowler.io
pulseData
Quanergy
Rainbird Technologies
RapidMiner
Ravn Systems
Re:infer
Realeyes
Recursion Pharmaceuticals
Reduced Energy Microsystems
Replika
Resnap
Retechnica
RobArt
RoboCV
Rokid
Sage Senses
Scaled Inference
Scortex
Seamless.AI
Seldon
Semantic Machines
SenseTime
Sensifai
Sentenai
Senter
Sentient
Sentisum
Sentrian
Shield AI
Sight Machine
SigOpt
Siwa
Skindroid
Skydio
Skymind
Sonalytic
SoundHound
SpaCy
SparkBeyond
SparkCognition
Speechmatics
Sportcaster
Tamr
Tend
Tenstorrent
TeraDeep
Terrabotics
Terraloupe
TheySay
ThinCI
Third Eye Systems
TickAI
Tractable
TUPU
Twenty Billion Neurons
TypeScore
Unisound
Universal Robotics
Valoosa
Velodyne
Vertex.ai
Vicarious
Visii
Visio Ingenii
Viz
Volley
Vuno
Wave Computing
White Matter
Xihelm
YITU Technology
Zephyr Health
Zero Labs
Zero Zero Robotics
Zest Finance
Zoox