How to start a deep
learning startup,
NOT from scratch
Mostapha Benhenda, Mindolia
Kyiv deep learning meetup,
13 september 2016
What is deep learning?
● Specialty of machine learning, which uses
'deep' neural networks, i.e. with many (>3)
layers.
● No need to really understand what is 'deep
learning' in order to use it, just apply it:
● Applied mostly to understand images, videos,
languages (text, DNA...) and speech.
Why starting a startup?
● No experience, no job ? Just hire yourself!
● Startup = easiest way to get a real job experience, with
an awesome boss: you!
● Acqui-hire >> hire
● Startup for ML beginners >>> Coursera, Kaggle
● ML for startup: easier than ML for big company (less
data, less optimization needed)
● Startup more difficult to start later: higher opportunity cost
(better job offers with experience): now or never!
How to start a DL startup: easy
Deep learning startup = startup using
deep learning. You need:
1. Idea
2. Team
3. Product using deep learning
4. Market
● These 4 things: done quickly, and in parallel
● Avoid perfectionnism!
● Improve the bottleneck, the weakest link
1. Elaborate an idea
● best idea: from your own problems
● In my case (facial recognition): ringing doorbell= noise pollution
● Focus on customer pain
● Don't think too much: idea is only a starting point
● No idea → clone other startups (see Angellist, Crunchbase...)
● See my list of 19 ideas:
https://docs.google.com/presentation/d/1Z-CPIGbSSTOm_EaqS5ks1V
...any questions?
2. Build a team
● Ideal team: 2 or 3 co-founders (Hipster + Hacker +
Hustler)
● Criteria of Minimal Viable Co-founders: trust,
motivation and skills
● No co-founder: start as a single founder
● Human co-founders disrupted by 'AI co-founders':
AWS, Google, Stackoverflow, Quora, blogs....
3. Assemble a deep learning
product
● Like IKEA: use ready-made parts
Minimal Viable Product (MVP):
● Design: sober and clean
● Code: quick and dirty
MVP= Deep learning+ Web app
Deep learning feature:
● Transfer learning (1 line of code+ little data)
● Open-source API: OpenFace, DeepDetect...
● Commercial API (Google, smaller companies...):
why not, but be careful of locking
● Don't start from scratch!!!
...any questions?
Web/mobile application
● Build your app locally first, then deploy
● Use LAMP: Linux Apache Mysql Python
● In my product, I used Twisted instead of Apache
because of live streaming
● Deployment: AWS or others (Microsoft, Google,
Heroku...)
● Debugging: use Google, Stackoverflow, and Rubber
Duck
4. Go to the market
● Code, technology: cheap moneypot
● Users, customers: valuable bees
Example: Uber Clone
● Cost: 2000 dollars with a freelancer
Original Uber
● Uber inc. is valued to 66 Billion dollars
Difference:
● Original Uber: 66 Million monthly trips
● Uber clone: zero trip.
● Conclusion: don't stop at coding, continue and find
users!!
Product/market fit
● Talk to potential users
● Monitor metrics, watch behavior
● Marketing, get visibility for your brand: communicate
with blogs: http://tinyurl.com/juy7exc
● Video clips:
https://www.youtube.com/watch?v=81btY-pjYeA
● ….any questions?
More advertising (for the meetup):
Hackathon 'Smart-techno' this weekend at Gulliver mall 24th
floor,
hands-on practice of this tutorial.
● Meetup agenda and suggestions: http://tinyurl.com/h5rl5ze
● Including 2 'orphan' Tensorflow tutorials, waiting for their instructors!
Adopt them, they are cute!
● Incentive: IF enough people study the tutorials VERY seriously (i.e. able
to give useful feedback),
THEN we will invite relevant experts for remote Q&A sessions!
…any questions?

Start a deep learning startup - tutorial

  • 1.
    How to starta deep learning startup, NOT from scratch Mostapha Benhenda, Mindolia Kyiv deep learning meetup, 13 september 2016
  • 2.
    What is deeplearning? ● Specialty of machine learning, which uses 'deep' neural networks, i.e. with many (>3) layers. ● No need to really understand what is 'deep learning' in order to use it, just apply it: ● Applied mostly to understand images, videos, languages (text, DNA...) and speech.
  • 3.
    Why starting astartup? ● No experience, no job ? Just hire yourself! ● Startup = easiest way to get a real job experience, with an awesome boss: you! ● Acqui-hire >> hire ● Startup for ML beginners >>> Coursera, Kaggle ● ML for startup: easier than ML for big company (less data, less optimization needed) ● Startup more difficult to start later: higher opportunity cost (better job offers with experience): now or never!
  • 4.
    How to starta DL startup: easy Deep learning startup = startup using deep learning. You need: 1. Idea 2. Team 3. Product using deep learning 4. Market
  • 5.
    ● These 4things: done quickly, and in parallel ● Avoid perfectionnism! ● Improve the bottleneck, the weakest link
  • 6.
    1. Elaborate anidea ● best idea: from your own problems ● In my case (facial recognition): ringing doorbell= noise pollution ● Focus on customer pain ● Don't think too much: idea is only a starting point ● No idea → clone other startups (see Angellist, Crunchbase...) ● See my list of 19 ideas: https://docs.google.com/presentation/d/1Z-CPIGbSSTOm_EaqS5ks1V ...any questions?
  • 7.
    2. Build ateam ● Ideal team: 2 or 3 co-founders (Hipster + Hacker + Hustler) ● Criteria of Minimal Viable Co-founders: trust, motivation and skills ● No co-founder: start as a single founder ● Human co-founders disrupted by 'AI co-founders': AWS, Google, Stackoverflow, Quora, blogs....
  • 8.
    3. Assemble adeep learning product ● Like IKEA: use ready-made parts
  • 9.
    Minimal Viable Product(MVP): ● Design: sober and clean ● Code: quick and dirty
  • 10.
    MVP= Deep learning+Web app Deep learning feature: ● Transfer learning (1 line of code+ little data) ● Open-source API: OpenFace, DeepDetect... ● Commercial API (Google, smaller companies...): why not, but be careful of locking ● Don't start from scratch!!! ...any questions?
  • 11.
    Web/mobile application ● Buildyour app locally first, then deploy ● Use LAMP: Linux Apache Mysql Python ● In my product, I used Twisted instead of Apache because of live streaming ● Deployment: AWS or others (Microsoft, Google, Heroku...) ● Debugging: use Google, Stackoverflow, and Rubber Duck
  • 13.
    4. Go tothe market ● Code, technology: cheap moneypot ● Users, customers: valuable bees
  • 14.
    Example: Uber Clone ●Cost: 2000 dollars with a freelancer
  • 15.
    Original Uber ● Uberinc. is valued to 66 Billion dollars
  • 16.
    Difference: ● Original Uber:66 Million monthly trips ● Uber clone: zero trip. ● Conclusion: don't stop at coding, continue and find users!!
  • 17.
    Product/market fit ● Talkto potential users ● Monitor metrics, watch behavior ● Marketing, get visibility for your brand: communicate with blogs: http://tinyurl.com/juy7exc ● Video clips: https://www.youtube.com/watch?v=81btY-pjYeA ● ….any questions?
  • 18.
    More advertising (forthe meetup): Hackathon 'Smart-techno' this weekend at Gulliver mall 24th floor, hands-on practice of this tutorial. ● Meetup agenda and suggestions: http://tinyurl.com/h5rl5ze ● Including 2 'orphan' Tensorflow tutorials, waiting for their instructors! Adopt them, they are cute! ● Incentive: IF enough people study the tutorials VERY seriously (i.e. able to give useful feedback), THEN we will invite relevant experts for remote Q&A sessions! …any questions?