Updated deeper overview of investor's look at machine learning / deep learning startups, with slight Russian accent. =)
Some slides are courtesy of Russia.ai and personally great friend @Petr Zhegin:
#23, #28 are from http://www.russia.ai/single-post/2016/09/21/Ten-Russian-speaking-venture-capital-funds-one-may-consider-to-back-an-AI-startup
#30 insights are from http://www.slideshare.net/RussiaAI/artificial-intelligence-investment-trends-and-applications-h1-2016
Victor Osyka of Almaz Capital, http://fb.com/victor.osika, http://medium.com/@victorosyka
2. Who is Victor
1
Been on both sides of the table: startup founder,
venture investor at US/Russia fund
www.almazcapital.com. On boards of Carprice,
StarWind, Nival, 2Can-iBox, Yaklass, RoboCV etc.
4 years in VC, 1 yr. startup co-founder, 3 yrs. in
consulting, engineering + LBS MBA edu.
Reach here osyka.victor@gmail.com
http://medium.com/@victorosyka I post here and
at facebook
www.linkedin.com/in/victorosyka
http://facebook.com/victor.osika
3. Google trends stats on AI/ML/DL/bigdata + deep learning patents
Technology: deep/machine learning helped a lot in many domains, more
progress to come
Portrait of a fundable startup is probably:
– May aim to taking some technology barrier
– Not hardware, agnostic if b2c/b2b, biz co-founder(s), creates barriers for entry,
arbitrages R&D cost by CIS geo, ideally HQ in SV
Globally, funding is steadily growing
Exits are done in Russia in 2016 even under sanctions: Itseez, Api.ai
Takeaways
2
14. Data
Computing
– Demand for data and computing power will increase even more as too
much data and power is required to slightly decrease the error rate in
models that power AI. Buy Nvidia stocks? =)
– AI moves into real time – e.g. live video analytics, driving
Progress in architectures
– Complexity of architectures will go up at the hardware and the neural
networks level
– # DL developers: 2.2K => 55K in 2016
# GPU developers globally: 120K in 2014 => 400K in 2016. (ML is also
done at GPUs, but many operate on non-DL stack)
Inflection point is a result of abundance of:
13
16. According to type of data used
Visual
Computer vision
– Online
– AR
– Offline: cameras, robotics, self-
driving, self-flying
Image processing
Sound
Voice/Music synthesis
– Deepmind’s WaveNet
Speech recognition
– One user
– Dialogues, team talks
Text
Auto-translation
Text processing / dialogues
Other
Control systems (reinforcement
learning)
Scientific problems
Current state of the art in machine learning
15
18. Visual
Computer vision
– Who will recognize faces more successfully
than 60-70% level of quality at scale more than
0.5-1M+ pics?
– Who will be able to identify various objects by
SKU?
– Breakthroughs to current state of health
images processing
– Extract meaning from content
Image processing
– Real-time video filters? Realtime AR?
Sound
Voice/Music synthesis
– Who will increase speed of WaveNet by factor
of 100-1000x so it would be usable in real life?
Speech recognition
– Who will recognize dialogues better than 50-
60% level of quality?
Text
Text processing / dialogues
– Chats with end user satisfaction of more than
20-30-40% Or more complex talks?
– Who can extract meaning?
Auto-translation
– What is better than google?
Other
Control systems (reinforcement learning)
– Who will do gaming better?
…and make autonomous agents based on
gaming spaces?
– Who will apply RL to other control domains
than power of servers etc.?
Scientific problems
– Any meaningful breakthroughs to current
states
What one should seek in technology?
17
20. Other than “purely product co” types of ML startups
19
Scientific company -
ML company with
radical tech
improvements
• Cross-disciplinary
team
• Aims to develop
new tech
• E.g. DeepMind,
Vicarious etc.
Research lab, not
company
• Develops new
knowledge
• And outsources it
• E.g. Open.ai, Caffe
library etc.
ML company with
incremental tech
improvements
• Inspired by others’
papers
• In house tech
optimization by
computer science
people
• Very clear product
focus
Product company,
productizing some
open sourced ML
tech stack but doing
very fast business
● e.g.: Prisma etc.
21. Software, not hardware
Doesn’t matter if b2c or b2b customers
– B2c good that scales virally if goes well +
uses crowdsourced data (see below)
– B2b is good that monetize-able +
accumulates proprietary data (see
below)
Many techs are replicated by followers
in 1-3 years, so business advantage
should be more complex
– Creates some barriers for entry/switch
costs.
e.g. acquire data either unique (e.g.
crowdsourced, not publicly
downloadable/parce-able), or at scale?
e.g. vertical market is targeted in a self-
reinforced data loop (more data = tech
performs better = customers are more
loyal). Example: health data, telco data,
industrial data, banking.
– Still, companies aiming to the taking
technology barriers are welcome
Team of not only tech ppl, some founder
must be product or biz obsessed
– Tech team can be big now, field seems to
be complex now
Exploits geo arbitrage for labor costs
– Gives more R&D headcount for the same
runway OR less $ needs to be raised
each time
Ideally, Russians in Valley: to be very
product/biz conscious by their living in
the ecosystem around + helps with next
rounds of fundraising
Portrait of an ideal fundable startup?
Confidential 20
22. Biz is critical. Sci/engineers problem is…
Confidential 21
Curiosity and freedom as a core value:
• “Disturbing me in my curious introspective
research”
• “Don’t touch me, boring biz guys.”
• “Customers are lamers” = no customer
listening, in essence`
• Market feedback is often perceived as an
annoying factor, limiting curiosity
23. Russian entrepreneurs
– Appear to miss some hot spots of the AI
landscape and focus their efforts on a
limited number of applications?
– Overly focused on consumer and robotics,
founders do not embrace cybersecurity,
finance and healthcare sectors, which are
considered to be among the hottest
themes…
Russian AI startups last few years
Confidential 22
Startup examples
Robotics
– Software
Toytemic, Krisaf, ExoAtlet
– Bots, drones, vehicles, DYI kits
Aeroxo, Endurance, Umki,
Sensepace, Wicron, OMI Plow,
Robodrom, Alpha Smart
Systems, xTurion, Anywalker,
Promobot, Bitronics Labs
Computer vision/Imaging
– 3DiVi, VisionLabs, CompVision,
Prisma, Life.Film, Vocord
Predictive analytics
– RCO, Medialogia, Eventos,
Promodern, Prometei,
Gloubhopper, Statsbot
AR/AR
– VRD, VR Systems, Kvadratik,
Bazelevs Innovation
Intelligent assistants
– Cubic, Findo, Lexy
28. By industry
Russian investors in foreign AI co’s (# = 38)
Confidential 27
By core technology
Altair – Youappi, Socure
Flint – CyberX, Youappi
Grishin Robotics – Occipital, RobotLab
I2BF – Planetary Resources, Autnomous
Marine Systems
LETA – Unomy, Visilights
Maxfield – Visilights, SpeakingPal
RTP – ReportGrid, WorkFusion
Runa – LendingRobot, TellmePlus, Digital
Genius
Titanium Investments – Feedviser, Mantis
Vision
Vaizra – PrimeSense, Face.com
29. Foreign investors
28
Top AI investors last 5 years
– Bloomberg Beta
Thesis is future of work/enterprise tech
– Google Ventures (series B-C-D, no seed or A)
– Samsung
Personal assistants and alike
– Rakuten
– Horizons
Assistants, text processing (e.g. ViV, made by founders of Siri)
– Intel Capital
Computer vision, hardware
– In-Q-Tel
– Khosla
Healthcare, general AI (Vicarious), ML platforms (Scaled Inference, Russian guy in USA)
30. In these 3 countries is the following industry breakdown of funded startups:
USA, UK and Toronto are AI clusters abroad?
29
34. Itseez by Intel, acquired in May 2016: 100 people in Nizhniy
Novgorod – sanctions does not matter if the target is so special
for the acquirer
Api.ai by Google, acquired in September 2016: also Russian
company
Exits in Russia still viable
33