AI for Decision Making Phanish Puranam, INSEAD - Lady Ada Lovelace
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
AI for Decision Making
Phanish Puranam, INSEAD
Lady Ada Lovelace
Phanish Puranam The Roland Berger Chaired Professor of Strategy and Organisation Design • PhD in Management from the Wharton School of the University of Pennsylvania • Served as the Academic Director for the PhD Program at both London Business School and INSEAD • Expertise: Organization Design & Strategy • Current research: Organizations & Algorithms, Remote Collaboration
Data Science
What does it take to master AI?
Computer Machine Math and
Science Learning Statistics
Unicorn
Traditional Traditional
Software Research
Subject Matter
Expertise
•
Copyright © 2014 by Steven Geringer Raleigh, NC. Permission is granted to use, distribute, or
modify this image, provided that this copyright notice remains intact
What’s the most impressive accomplishment of AI to date (in your view)? • Please enter into our Shared Google Doc- do check for redundancy with what others are typing! • 3 minutes
What’s the most impressive accomplishment of AI to date (in your view)?
Can AI be creative?
1959
1997
2016
Herbert Simon (joint work with W.G. Chase, 1973)
HUMAN EXPERTISE= INTELLIGENT BEHAVIOR IN A CONTEXT= PATTERN RECOGNITION !
OLD AI: EXECUTES RULES NEW AI: LEARNS PATTERNS FROM DATA
Pattern recognition & application • Learning to recognize and apply patterns is the heart of intelligence • That’s what Machine Learning does • Lots of data + processing power makes it easier to recognize (even complex) patterns • A pattern= a “function” – which associates inputs to unique outputs. • Pattern recognition in humans = function approximation in algorithms
A STORY OF DOGS AND CATS Supervised & Unsupervised Learning Reinforcement Learning
ML algorithms work through pattern
recognition & completion
Pattern Recognition &
(Unsupervised) Completion
(Reinforcement)
(i.e. function estimation
and application) Adaptation
Categorization
Feedback
Do X1 → F1
X1 X2. X3 X4. XN
Do X2 →F2
(Supervised) Do X2 → F3
A B ? Do X4 →F4
Do ?
Prediction
X1-- Y1
X2--Y2
XN-- ?Patterns → Predictions → Decisions • All decisions require prediction • Decisions can be “standalone” vs. “embedded”
AI ~ ML~ Pattern Detection: Standard use
cases today
Transaction features ---- Fraud
Customer attributes---- Churn likelihood
Inkdots ---- WordsAI ~ ML~ Pattern Detection: Standard use
cases today
Pixel ----- Image
Sound ------ Words in speech
https://www.youtube.com/watch?v=D5VN56jQMWM&t=127s
State of play ----- Optimal next move in
game
https://www.eff.org/ai/metrics#Abstract-Strategy-GamesAI ~ ML~ Pattern Detection : Emerging Use
cases
Employee attributes----Exit likelihood
CV attributes ---- Good hire
Project characteristics ---- Good investmentAI ~ ML~ Pattern Detection : Emerging Use
cases
Text and word use ----- Cultural values
Company characteristics----- Strategic movesHans Moravec’s landscape, in Tegmark (2018)
Dominance or Co-existence?
37
78HOW TO KEEP UPTODATE ON WHAT AI IS DOING (WELL)
https://www.eff.org/ai/metricsLevels of sophistication in using Data in Business
PROTOTYPING
PREDICTION
PERCEPTIONY
Data
Continuous
Outcome
•
X
@Phanish PuranamPERCEPTION vs. PREDICTION
• Perception through • Prediction through
Hypothesis testing: Data mining:
• Does this association • What associations
exist in the data? exist in the data?
• We have a hunch (hypothesis)
about an explanation • We have no or minimal hunches,
• Null Hypothesis Significant we “let the data speak” to make
Testing tells us the probability a prediction
(across many samplings) of • Data mining involves algorithms
observing the association we see that hunt for useful associations
merely by chance, even if the in the data that can be
true association in the summarized in a model.
population is zero.
• p-value • Predictive accuracyPrediction What’s going to happen?
•
x ? y
• Different algorithms • Common to all:
• Tree induction • OLS
• Use past data to predict
• LASSO • Logistic
• future outcomes
• Random forest • Deep learningHow can more advanced ML algorithms improve on what we just did?
Y
Past Data
Continuous
Outcome
Key issue: build models
that fit current data well
but also predict future
data well
•
X
@Phanish PuranamOverfitting + Generalization • Problem: • Solutions: • Every sample is 1. Penalizing predictive accuracy for model truth + error. The complexity model may learn 1. Regularization (constraint on parameters) too much about 2. Partition training data into training and test data the error in a 1. K-fold validation particular sample by adding too 3. Both! many parameters.
Components of a prediction machine
Data Algorithms Predictions
1.Representative 1. Produces highest
2.Captures structure predictive accuracy
3. Stable 2. At least sacrifice of That have high value-
4.Free of social biases interpretability accuracy slope
https://knowledge.insead.edu/blog/insead-blog/where-ai-can-help-your-business-and-where-it-cant-13136Performance
Type “A” tasks Type “B” tasks
Division of Labour + Integration of Effort
Human
Human + Algorithm
AlgorithmPerformance
Type “A” tasks Type “B” tasks Type “C” tasks
Division of Labour + Integration of Effort Wisdom through Aggregation
Human
Human + Algorithm
AlgorithmYou can also read
