FOUR RULES: (1) AI is a tool not a business model. (2) Protect your data but use federated learning to share models. (3) Regulation should guide you & not stop you - use tools like LIME. (4) Think holistically and build a data culture of fair data usage.
8. LutzFinger.comLutzFinger.com
LutzFinger
2nd. Wave: More Rules
Source: Wikipedia
BEFORE:
● AI limited to small
applications like
Checkers
NOW:
● Aiming to build
systems to augment
the knowledge of
doctors
“We know more than we can tell”
Michael Polany 1958
13. LutzFinger.comLutzFinger.com
LutzFinger
AI Can
Create New Products
https://eng.uber.com/atg-dataviz/
https://www.symrise.com/newsroom/article/breaking-new-fragra
nce-ground-with-artificial-intelligence-ai-ibm-research-and-symr
ise-are-workin/
New Products Can
Create New Business
16. LutzFinger.comLutzFinger.com
LutzFinger
The Levers
• Direct Relationship with Users
• They Own the Relationship
• No One is between Demand and them
• Zero Marginal Costs For Serving Incremental Users
• No COGS
• No Distribution Cost
• No Transaction Cost
• Abundant Supply
• With either zero marginal acquisition cost
• OR acquisition cost that scales
Those are attractive terms for the business. Data improves over time
the service and creates a moat for competition. But data does not
create the business model.
Further reading: https://stratechery.com/
18. LutzFinger.comLutzFinger.com
LutzFinger
Tools Create New Product & New Business
Tools Help To Optimize Your Core Biz.
https://eng.uber.com/atg-dataviz/
https://www.symrise.com/newsroom/article/breaking-new-fragra
nce-ground-with-artificial-intelligence-ai-ibm-research-and-symr
ise-are-workin/
● Banking: Card Fraud Detection
● Banking: Credit Scoring
● Media: Content Recommendation
● Health Care: Fraud Detection
● Medicine: Image Processing
● Medicine: Outliers Detection
● Retail: Likelihood to Buy
● Books: Marketing Planning
● Manufacturing: Failure Prediction
● Manufacturing: Optimization
● Insurance: Likelihood & Pricing
● Transportation: Route Planning
● Energy: Grid Utilization
19. LutzFinger.comLutzFinger.com
LutzFinger
Insurance
● Damage Estimations: Using Deep Learning &
Satellite Images to understand damage of the
Camp Fire in 2018
● Insurance Premiums:
Praedicat mines
literature to extract
metadata about bodily
harm that can be
attributed to certain
materials.
https://www.space.com/41403-california-wildfires-2018-photos
-from-space.html
● Claim Management & Indemnity Prediction: Digital
workflows improve workflows and predictions (AI
Insurance)
20. LutzFinger.comLutzFinger.com
LutzFinger
Oil & Gas
● Clean up data from pressure transient analysis (PTA)
using Singular Spectrum Analysis (SSA) and a
supervised learning
● Methode can replace the manual task of identifying the
flow regime.
https://doi.org/10.1145/3292500.3330661
https://www.kdd.org/kdd2019/accepted-papers/view/structured-noise-detection-application-on-well-test-pressure-derivative-dat
21. LutzFinger.comLutzFinger.com
LutzFinger
Retail Credit Loan Applications
(RNN over debit or a credit card transactions)
1) More accurate than
traditional
2) Less human overhead (raw
transaction data >> no
feature needed)
3) Fast as no additional
customer input is needed
(transactional data)
4) Hard to forge signals
(transactional data)
5) Works for customers without
credit history https://doi.org/10.1145/3292500.3330693
23. LutzFinger.comLutzFinger.com
LutzFinger
Fitbit for Cows
Helping to manage health from cows to improve milk
production.
The only difference between cows that produce 30
liters of milk a day and those that produce 10 liters
was the animal’s health.
https://itunes.apple.com/app/ida-by-connecterra/id1246795438
https://www.connecterra.io/ida/registration-for-ida-event-at-neu-hope-dairy-usa/ida-dashboard-en/
36. LutzFinger.comLutzFinger.com
LutzFinger
Transfer Learning at Work
Combining historical mining and exploration data with NASA data and use pre-trained deep
learning networks to identify several target zones with the highest potential for gold
mineralization in Nevada’s Jerritt Canyon district.
39. LutzFinger.comLutzFinger.com
LutzFinger
Regulation Considerations
Ability to Explanation
“[...] the controller shall provide [...]
the following information: the
existence of the automated
decision-making and [...] meaningful
information about the logic involved
and the envisaged consequences of
such processing for the data
subject.”
Art. 13-15 & 22 Regulation (EU)
2016/679
40. LutzFinger.comLutzFinger.com
LutzFinger
One solutions: LIME
( local interpretable model-agnostic explanations)
https://github.com/marcotcr/lime AND https://homes.cs.washington.edu/~marcotcr/blog/lime/
Predict on PermutationTake Data Point(s) = Local Permute Data Points (+ Noise)
Use distance and other similarity
metrics to understand the most
important (not permuted) features.
48. LutzFinger.comLutzFinger.com
LutzFinger
Your Cookbook
1. AI is a Tool Not A Business Model
a. Not everything is a unicorn
b. Biz improvements are a good start
2. Not More Data But “Right” & “Protected” Data
a. Own / protect the access
b. Use federated learning to share models
c. There are diminishing returns in data
3. Regulation Should Guide You Not Stop You
a. Use LIME to understand the model
b. Test for Biases
4. Think Holistically
a. Hire data scientists and engineers
b. Set a AI minded culture