3. 3
FROM THINKING MACHINES… TO DETECTING CATS AND DOGS
MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 2018
4. 4
MACHINE LEARNING: RESEARCH & INDUSTRY
MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 2018
5. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 20185
ZALANDO
• 17 Countries
• 7 locations in Europe
• 4.5 bn € revenue 2017
• ~15.000 employees
• IPO Oct 2014
https://geschaeftsbericht.zalando.de/2017/geschaeftsbericht/zahlen-daten-und-fakten/
6. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 20186
SEARCH PLATFORM
7. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 20187
WAS I RIGHT?
8. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 20188
WAS IT WHAT I EXPECTED?
🤔
10. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201810
• What is out there: AI Hierarchy of needs + methods + tools = AI?
SO HOW DO YOU “DO AI?”
https://hackernoon.com/the-ai-hierarchy-of-needs-18f111fcc007
11. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201811
ORGANIZATION OF WORK
12. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201812
HOW?
14. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201814
• Customer has a Job to be
done.
• Business has to give a
Solution that adds value.
• Solution consists of
Activities, and some of that
activities might be solved
with ML.
• ML: solve a task by learning
from examples.
FROM BUSINESS PROBLEMS TO MACHINE LEARNING PROBLEMS
15. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201815
MACHINE LEARNING: LEARN FROM EXAMPLES
16. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201816
EXAMPLE: RECOMMENDATIONS
17. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201817
Typical problem for machine learning:
• Hard to specify what exactly means “similar.”
• A lot of example data is available.
• Recommendations have to change based on new articles
frequently.
EXAMPLE: RECOMMENDATION
18. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201818
Typical algorithms:
• Collaborative
filtering,
• Content based
recommendation,
• Predicting customer‘s
next action.
EXAMPLE: RECOMMENDATIONS
20. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201820
Data Scientists and Developers
21. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201821
THE SCIENTIFIC METHOD
22. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201822
Very different approaches to
coding…
← developers
data scientists →
DS&D: Coding
23. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201823
• What is the most
productive way?
• Ideally, interface on
code, not just
documentation
• Production logs often
become data analysis
input!
DS&D: Collaboration
24. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201824
ADDING ML TO THE MIX
25. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201825
ADDING ML TO THE MIX
26. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201826
ORGANIZING DATA SCIENCE
32. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201832
33. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201833
MARVIN MINSKY: RECOGNIZING SIMPLE PICTURES
http://web.media.mit.edu/~minsky/papers/PR1971.html
34. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201834
AI TIMELINE PAST 20 YEARS (that’s 1998 till 2018)
35. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201835
• Initially approached AI like any
other problem computers could
solve.
• Alternatively, using an approach
inspired by human biology.
• Machine Learning added a
statistical approach to the mix.
• Recently, Deep Learning has led to
impressive improvements.
APPROACHES IN ARTIFICIAL INTELLIGENCE
36. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201836
• Classical approach is to specify
what the input/output relation is,
then devise programs to solve
that.
• Machine Learning replaces that
with examples (+ a cost
function).
• Training then means to infer a
model that generalizes well on
future data.
BIRD’S EYE VIEW OF MACHINE LEARNING
37. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201837
• Artificial Intelligence is the overarching goal or challenge.
• Machine Learning is one approach that has proven very successful if the problem
itself cannot be specified easily.
ARTIFICIAL INTELLIGENCE VS. MACHINE LEARNING
38. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201838
• Many reasons why:
convenience, security,
disrupting mobility.
• Current approaches are a mix
of many systems, some of
which make heavy use of
machine learning.
• Deep Learning very successful
for computer vision and image
analysis.
AUTONOMOUS DRIVING
39. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201839
Autonomous driving is based on a mix of sensors with quite
different capabilities to improve reliability.
• Sonar/Radar: Cheap, low resolution, works well under
extreme weather and in darkness, can estimate velocity.
• Camera: Cheap, very high resolution, similar to what we
humans see.
• Lidar (light detection and ranging): expensive, very
accurate depth maps.
SENSORS IN AUTONOMOUS DRIVING
40. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201840
• Companies like Waymo do extensive data
collection and simulation to evaluate and tune
the system.
• Not just for training ML methods, also for overall
systems testing.
ML inspired approach to defining the problem, but mix
of ML and explicit solutions.
(Waymo lecture at MIT)
DATA-DRIVEN APPROACH TO AUTONOMOUS DRIVING
41. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 2018
• Not that much ML in
there (at least right
now).
• Dialog is done through
Frames that capture a
piece of information
required and an
analysis part that maps
user input to fields.
• ML used for
understanding
speech2text, named
entity recognition,
analysis
41
CHATBOTS
Example 1:
A: “I’d like to book a flight tomorrow”
B: “From where to where do you want to fly?”
A: “From London to Berlin.”
B: “With how many passengers?”
A: “Just me.”
B: “Okay, so I have one passenger for a flight
from London to Berlin tomorrow. Is that
correct?”
A: “Yes.”
B: …
Example 2:
A: “I’d like to book a flight for me
tomorrow from London to Berlin.”
B: …
Booking a flight:
Frame:
- when: Date
- start, end: Location
- how many persons: Number
42. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201842
• Machine Learning
used especially on
“perception” part.
• Core is rule based
system.
• Potential to improve
those based on
examples, too, same
for text2speech.
CHATBOT SYSTEM OVERVIEW
43. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201843
Recommendations as an AI problem:
• Understand what the user is
looking for right now. What is his
intent, what is in his mind?
• Technically, predict next action.
• Quite involved, dealing with real-
time data, etc.
RECOMMENDATIONS
44. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201844
• From computer science’s point
of view, strategy games are
“easy” if you know the value of
each state.
• Cleverly simulating “plausible
actions” leads to speedup
(Monte Carlo tree search)
ALPHA GO AND OTHER STRATEGY GAMES
45. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201845
ALPHA GO: CONVOLUTIONAL NEURAL NETWORKS FOR POLICY AND VALUE PREDICTION
46. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201846
• Machine Learning used as part of the system
• Otherwise lots of engineering and “classical approaches”
• Main difference IMHO is “data driven” vs. “spec driven” approach.
SOME MODERN “AI” SYSTEMS
47. MIKIO BRAUN: “ACADEMIA TO INDUSTRY: LOOKING BACK ON A DECADE OF AI” IDALAB SEMINARS #14, BERLIN, NOV 2, 201847
• Image recognition almost on human level performance.
• Outperforms humans on strategy games.
• Able to learn features by itself.
• Can act intelligently and autonomously in the real world.
YES!
48. What is Deep Learning - StrataData New York, 2017 - Mikio Braun48
• From raw data to symbolic representations?
• Capability to self reflect?
• What happens inside the machine? Is it just imitating?
• Learning from single examples.
NO!
50. This presentation and its contents are strictly confidential. It may not, in
whole or in part, be reproduced, redistributed, published or passed on to
any other person by the recipient.
The information in this presentation has not been independently verified. No
representation or warranty, express or implied, is made as to the accuracy
or completeness of the presentation and the information contained herein
and no reliance should be placed on such information. No responsibility is
accepted for any liability for any loss howsoever arising, directly or
indirectly, from this presentation or its contents.
DISCLAIMER
50
Editor's Notes
- Always interested in AI and Model of the Mind.
- As soon as I could => ML
Just about prediction
But that is actually great
=> what is the core of ML?
- Name
- Position
- Story of Academia to Industry and what I learned about ML
Fulfillment center == warehouse
135M visits/month =~ 450K visits/day
Active Customer: Anyone who has placed an order in the past 12 months
⅓ of 9000 in Berlin
Spend some time
Explore
Explore islands, look for loops
Global system to facilitate: peer reviewed publications focussed on novelty
=> something is missing.
Startup on the side
Pilot projects
Two engs, didn’t take off => industry
ML at the core
Layer of infrastructure to deliver product
Automate what was one off
Research still there, at product level
=> so how do we compare?
research explores / industry builds
Simple stuff works best, because research is often exploring
Similar to toolsmiths, but overlooked because everything is software
=> Leaps: Deep Learning
This is eniac. Ever since people built the first computers, the idea was to build a machine that thinks, to understand the mechanics of mental work, but we didn’t really know what this is.
Famously, Marvin Minsky in 1966 posed this as a summer project, and years later it wasn’t solved.