AI Fundamentals for Non-Data Scientists

AI for Business Introduction

Kartik Hosanagar, Professor of Operations, Information and Decisions

AI and Business

• Artificial intelligence (AI) is about getting computers to do things that

require human intelligence
• Understanding language
• Reasoning
• Navigating the physical world
• Learning
• Machine learning - focused on getting computers to learn without
explicitly programming them
• AI is increasingly being seen as the next phase of digital transformation
Digital Transformation Timeline

• Multiple digital technologies have helped transform business

• Companies that were slow to react paid the price
Internet Cloud Computing Mobile Computing AI

Late 1990s Mid 2000s Late 2000s Now

• AI is likely to be equally transformative and can potentially be viewed as a

general-purpose technology
What are General-Purpose Technologies (GPTs)?

• GPTs are “characterized by the potential for pervasive use in a wide

range of sectors and by their technological dynamism” (Bresnahan &
Trajtenberg)
• GPTs stimulate innovation and economic growth
• GPTs inform products strategy & organizational design/strategy

• 3 factors can indicate that a new technology shows promise as a GPT:

1.Widespread use of the technology across many industries
2.High volume of research jobs related to the technology
3.Research jobs widespread across many industries
Content/quotes from: “Could Machine Learning Be a General-Purpose Technology? Evidence from Online Job Postings” by Goldfarb, Taska & Teodoridis.
Note: the Bresnahan & Tajtenberg quote is from the above paper. The authors of that paper had quoted this definition from Bresnahan & Tajtenberg
How Does AI Show Promise as a GPT?
Research by Goldfarb et al (2019)

Total Percent
Technology Total Jobs Research Research
Jobs Jobs

Machine Learning 370,572 5715000% 14.6%

GIS Software 142,645 899300% 5.8%

High volume
CRISPR 6,716 525000% 78.4%
•
of jobs and
Research jobs are a particularly
Quantum Computing 6,355 134800% 17.4%
research
important jobs
indicator of GPTs
Fracking 2,957 14100% 2.8% because they demonstrate that
Robotics 284,474 2244800% 7.3% a technology is “capable of
ongoing improvement” and
Nanotechnology 4,883 251900% 46%
could have significant future
Internet-of-things 124,506 575000% 4.3% potential, beyond what is
currently recognized
Cloud Computing 1,415,289 2939200% 2.2%

Content/quotes from: “Could Machine Learning Be a General-Purpose Technology? Evidence from Online Job Postings” by Goldfarb, Taska & Teodoridis.
ML’s Promise as a GPT (cont.)
Jobs widespread across industries
Industry NAICS2 AI GIS Quantum Fracking Robotics Nanotech IoT CRISPR Cloud Total
Accommodation and Food Services 5,389 883 4 5 1,516 5 828 2 15,090 5,594,113
Admin & Support & Waste Mgmt & Remediation Srvcs. 4,142 2,099 0 75 2,363 6 1,836 9 26,763 2,237,225
Agriculture, Forestry, Fishing and Hunting 820 1,388 0 3 384 0 146 0 2,310 97,644
Arts, Entertainment and Recreation 437 917 0 2 925 7 109 0 2,591 578,889
Construction
Educational Services
335
12,467
2,406
6,836
1
246
45
62
2,330
10,868
22
1,764
361
1,503
High volume
0
1,384
3,709
17,848
850,081
4,018,454

of jobs and
Finance and Insurance 40,261 3,871 26 32 7,809 7 4,569 35 98,700 5,787,861
Health Care and Social Assistance 102,401 34,512 158 612 114,965 734 28,477 1,793 481,770 38,700,000
Information 32,613 4,483 189 26 6,213 85 17,079 2 130,353 2,229,564
Management of Companies and Enterprises
Manufacturing
348
37,215
302
7,738
0
257
4
144
160
73,053
2
582
108
24,472
research jobs
0
1,733
1,156
96,368
80,359
6,307,396
Mining 674 1,098 0 828 489 2 257 0 1,736 216,880
Other Services (except Public Administration) 1,614 1,695 0 77 2,314 12 181 5 5,786 1,120,323
Professional, Scientific, & Technical Services 97,704 39,478 5,403 395 36,715 1,550 33,912 1,683 435,035 7,333,834
Public Administration 4,838 21,175 41 29 3,467 70 680 46 13,446 2,448,184
Real Estate Rental and Leasing 2,086 3,302 5 16 3,815 0 500 1 8,000 1,293,048
Retail Trade 21,705 1,220 14 132 11,214 21 5,917 2 54,247 8,720,486
Transportation and Warehousing 3,411 2,639 1 398 2,803 6 1,236 2 11,433 4,668,538
Utilities 801 3,896 0 51 886 6 493 1 2,877 333,522
Wholesale Trade 1,311 2,680 10 21 2,185 2 1,842 18 6,071 620,793
Total 370,572 142,645 6,355 2,957 284,474 4,883 124,506 6,716 1,415,289 93,237,194

Table A3.1. Number of jobs in data by industry and technology (2015-2018) Research by Goldfarb et al (2019)
ML’s Promise as a GPT (cont.)
Jobs widespread across industries
Industry NAICS2 AI GIS Quantum Fracking Robotics Nanotech IoT CRISPR Cloud Total
Accommodation and Food Services 5,389 883 4 5 1,516 5 828 2 15,090 5,594,113
Admin & Support & Waste Mgmt & Remediation Srvcs. 4,142 2,099 0 75 2,363 6 1,836 9 26,763 2,237,225
Agriculture, Forestry, Fishing and Hunting 820 1,388 0 3 384 0 146 0 2,310 97,644
Arts, Entertainment and Recreation 437 917 0 2 925 7 109 0 2,591 578,889
Construction
Educational Services
335
12,467
2,406
6,836
1
246
45
62
2,330
10,868
22
1,764
361
1,503
High volume
0
1,384
3,709
17,848
850,081
4,018,454

of jobs and
Finance and Insurance 40,261 3,871 26 32 7,809 7 4,569 35 98,700 5,787,861
Health Care and Social Assistance 102,401 34,512 158 612 114,965 734 28,477 1,793 481,770 38,700,000
Information 32,613 4,483 189 26 6,213 85 17,079 2 130,353 2,229,564
Management of Companies and Enterprises
Manufacturing
348
37,215
302
7,738
0
257
4
144
160
73,053
2
582
108
24,472
research jobs
0
1,733
1,156
96,368
80,359
6,307,396
Mining 674 1,098 0 828 489 2 257 0 1,736 216,880
Other Services (except Public Administration) 1,614 1,695 0 77 2,314 12 181 5 5,786 1,120,323
Professional, Scientific, & Technical Services 97,704 39,478 5,403 395 36,715 1,550 33,912 1,683 435,035 7,333,834
Public Administration 4,838 21,175 41 29 3,467 70 680 46 13,446 2,448,184
Real Estate Rental and Leasing 2,086 3,302 5 16 3,815 0 500 1 8,000 1,293,048
Retail Trade 21,705 1,220 14 132 11,214 21 5,917 2 54,247 8,720,486
Transportation and Warehousing 3,411 2,639 1 398 2,803 6 1,236 2 11,433 4,668,538
Utilities 801 3,896 0 51 886 6 493 1 2,877 333,522
Wholesale Trade 1,311 2,680 10 21 2,185 2 1,842 18 6,071 620,793
Total 370,572 142,645 6,355 2,957 284,474 4,883 124,506 6,716 1,415,289 93,237,194

Table A3.1. Number of jobs in data by industry and technology (2015-2018) Research by Goldfarb et al (2019)
ML’s Promise as a GPT (cont.)
Research jobs also widespread across industries
Industry NAICS2 AI GIS Quantum Fracking Robotics Nanotech IoT CRISPR Cloud Total
Accommodation and Food Services 438 24 1 0 43 2 32 2 214 13,410
Admin & Support & Waste Mgmt & Remediation Srvcs. 506 56 0 3 69 1 55 9 472 16,768
Agriculture, Forestry, Fishing and Hunting 243 220 0 0 24 0 6 0 130 9,847
Arts, Entertainment and Recreation 51 110 0 0 12 4 10 0 57 7,823
Construction
Educational Services
53
3,991
76
1,785
0
71
0
14
68
1,979
1
682
13
139
High volume
0
1,095
69
1,023
6,204
267,403

of jobs and
Finance and Insurance 6,409 271 15 3 309 2 249 29 2,132 114,244
Health Care and Social Assistance 11,974 1,905 63 21 8,186 366 895 1,393 7,299 1,078,322
Information 6,148 144 19 0 584 21 404 1 3,383 53,490
Management of Companies and Enterprises
Manufacturing
42
7,682
7
382
0
103
0
35
6
6,024
0
248
9
1,598
research jobs
0
1,363
19
3,306
1,061
357,069
Mining 108 51 0 25 34 1 70 0 37 4,251
Other Services (except Public Administration) 273 162 0 3 102 2 3 0 104 15,440
Professional, Scientific, & Technical Services 13,595 1,423 1,048 23 3,704 1,134 1,798 1,305 8,845 391,943
Public Administration 728 1,544 17 5 401 33 55 37 287 149,862
Real Estate Rental and Leasing 371 544 2 1 83 0 30 0 193 9,976
Retail Trade 3,599 129 9 2 570 14 290 2 1,470 35,200
Transportation and Warehousing 643 98 0 1 95 1 27 1 188 13,639
Utilities 99 28 0 5 29 5 19 0 36 5,762
Wholesale Trade 197 34 0 0 126 2 48 13 128 16,492
Total 57,150 8,993 1,348 141 22,448 2,519 5,750 5,250 29,392 2,568,206

Table A3.2. Number of research jobs in data by industry and technology (2015-2018) Research by Goldfarb et al (2019)
AI and Business
• AI (specifically ML) shows early indicators of being a general-
purpose technology
• Implications for firms if ML is a GPT:
• Realize that most industries are likely to change
• Be patient: the transformative impact may come with a lag
• To effectively leverage the opportunities, managers need to
understand the technology and its applications, and make
changes to their “business models, technology infrastructure,
organizational processes, and culture”
Content from: “Could Machine Learning Be a General-Purpose Technology? Evidence from Online Job Postings” by Goldfarb, Taska &
Teodoridis
Content/Quote from: https://www.simplilearn.com/introduction-to-digital-transformation-tutorial
AI Fundamentals for Non-Data Scientists
Course Introduction

Kartik Hosanagar, Professor of Operations, Information and Decisions

Course Introduction

This course will discuss AI Fundamentals from a business perspective

• Introduction to Big Data
• Introduction to Artificial Intelligence (AI)
• Machine Learning (ML) in Practice
• Building AI Systems Without Data
AI Fundamentals for Non-Data Scientists
Big Data Overview

Kartik Hosanagar, Professor of Operations, Information and Decisions

Agenda for Module 1

• Big Data Overview

• What big data is, how it is being generated, and why it matters

• Big Data Analysis

• Approach to analysis, analytics competencies, and broad skillset needs

• Big Data Tools

• Data management tools and data analysis tools

• Extracting Intelligence from Big Data

• Predictive analytics and implications for business strategy
“Data is the new Oil. Data is just like crude. It’s
valuable, but if unrefined it cannot really be used.”
— CLIVE HUMBY
“ We have for the first time an economy based on a key resource
[Information] that is not only renewable, but self-generating.
Running out of it is not a problem, but drowning in it is.”
— JOHN NAISBITT
What is Big Data?

• Data has been very important to businesses for multiple decades, but the
emphasis on big data is relatively new

• Data that “exceeds the capacity or capability of…conventional methods and

systems” (National Institute of Standards and Technology)

• Big data is not only about volume of data, it is also about:

• The structure of the data set
• The speed at which it is created
• The tools you need to analyze it
• What you can do with the data set

Quote from: https://www.technologyreview.com/s/519851/the-big-data-conundrum-how-to-define-it/

Additional content from: “Big Data” by Victor Mayer- Schonberger  and Kenneth Cukier  and https://www.sas.com/en_us/insights/big-data/what-is-big-data.html
Big Data Characteristics

New data characteristics created by today’s digitized marketplace

“Terabytes to petabytes “Structured,

of existing data to unstructured, text,
process” multimedia”

“Uncertainty due to data

inconsistency &
“Streaming data,
incompleteness,
milliseconds to seconds
ambiguities, latency,
to respond"
deception, model
approximations”

Image from: https://www.cbpr.me/big-data-and-public-relations/4vs/

Additional quotes from another diagram found at https://www.datasciencecentral.com/profiles/blogs/data-veracity
Drivers of Big Data

• Computing capacity: the capacity to store data has increased and the
associated cost has decreased

• Data generation: the world is going digital, more people and things are
connected than ever before

Content/quotes from: “Big Data” by Victor Mayer- Schonberger  and Kenneth Cukier, Images: https://www.axios.com/fewer-people-use-facebook-for-news-
13273e71-ab51-49cc-b4ee-1588a241c0c8.html, https://www.imore.com/best-fitbit, https://www.digitaltrends.com/cell-phone-reviews/apple-iphone-xs-review/
Drivers of Big Data

Big data allows you to:

1. Ask new questions 2. Answer same questions better

• This can be done across industries

• Healthcare, education, transportation, and more

Content/quotes from: “Big Data” by Victor Mayer- Schonberger  and Kenneth Cukier
 Big Data in Healthcare

Big Data Result

• Data about heart • Improves
rate, sleep quality, tracking of
exercise and more health patterns
• Gathered
continuously and
available in real
time

Availability

Content/quote from: https://intellipaat.com/blog/7-big-data-examples-application-of-big-data-in-real-life/, Images: https://support.apple.com/guide/watch/heart-rate-apda88aefe4c/watchos,

https://www.theverge.com/2017/11/30/16719458/apple-watch-study-irregular-heart-rhythms-stanford-university, https://www.wareable.com/apple/apple-watch-activity-and-workout-app-explained-875
 Big Data in Transportation

Big Data Result

• Data about traffic, • Allows for better
road closures, route planning
accidents, etc. and scheduling
• Available in a
visual format in
real time

Availability

Content/quote from: https://intellipaat.com/blog/7-big-data-examples-application-of-big-data-in-real-life/, Images: https://www.nytimes.com/2019/01/21/upshot/stuck-and-stressed-the-health-costs-of-

traffic.html, https://medium.com/@imtechpros_87395/where-does-google-maps-get-its-traffic-data-from-2562f984d82f, https://www.developerlibs.com/2018/08/flutter-how-can-draw-route-on-google.html
AI Fundamentals for Non-Data Scientists
Big Data Analysis

Kartik Hosanagar, Professor of Operations, Information and Decisions

Big Data Approach to Analysis

Traditional Analytics Big Data Analytics

Structured & Repeatable Iterative & Exploratory

Question Hypothesis Data Exploration

Answer Data Actionable Insight Correlation

Start with hypothesis Data leads the way

Test against selected data Explore all data, identify correlations
Big Data Requires a Broad Skillset

Manage the data

Tool Developers
Data Experts: Data architects, governance, policy

Understand the data

Data Science: Statistics, computer science

Visualization Expertise: Interpret data, graph them in meaningful ways

Act on the data

Decision Making - Exec. & Management: Apply data to solve business issues
Industry Vertical Domain Expertise: Identify relevant business issues, ask the right questions
Choosing a Big Data Tool

• Two broad categories of big data tools to choose from, depending on

whether you are trying to manage the data or analyze it

Data Management Tools Data Analysis Tools

“Data Mining”

Data Warehouses Clustering

Hadoop & Spark Association Rule Mining

Machine Learning
AI Fundamentals for Non-Data Scientists
Data Management Tools

Kartik Hosanagar, Professor of Operations, Information and Decisions

What is a Data Warehouse?

• Database

• A structured collection of data

• Example: an Excel spreadsheet
• Database Management System (DBMS)
• Allow users to access and manage the database
• Example: Excel, Microsoft, Oracle, MongoDB
• Often referred to as a database
What is a Data Warehouse?

• Data warehouse
• A particular kind of DBMS
• Specialized in the type of data it stores - historic data from many sources
• Specialized in the purpose it serves - analytics
• Example: Azure SQL Data Warehouse, Google BigQuery, Snowflake,
Amazon Redshift

Content/quotes from: https://www.dremio.com/what-is-a-data-

warehouse/
 How Do Data Warehouses Work?

Diagram from: https://www.dremio.com/what-is-a-data-warehouse/

Value of a Data Warehouse

• A single point of access for data

• A history of all the data is stored
• Separates operations from analytics
• Allows answers to questions like: “How much revenue has each product line
brought in per month over the last 10 years, broken out by month, city, and
state?”

All information quoted from: https://www.dremio.com/what-is-a-data-warehouse/

Open Source Big Data Tools: Hadoop and Spark

• Big data tools like Hadoop serve two main purposes, storage and processing
• Storage - lots of data (usually in a distributed fashion)
• Processing - involves distributed processing across nodes, parallelizing
the data process as much as possible
• Popular Hadoop distribution: Cloudera

Content (on 4 steps of Hadoop w/ MapReduce) from: https://hci.stanford.edu/courses/cs448g/a2/files/map_reduce_tutorial.pdf

Open Source Big Data Tools: Hadoop and Spark

• Spark is a more recent version of Hadoop

• Databricks is the dominant company built around Spark

Content (on 4 steps of Hadoop w/ MapReduce) from: https://hci.stanford.edu/courses/cs448g/a2/files/map_reduce_tutorial.pdf

AI Fundamentals for Non-Data Scientists
Data Management Infrastructure

Kartik Hosanagar, Professor of Operations, Information and Decisions

What is a Data Warehouse?

• Database:
• “A structured collection of data”
• Example: an Excel spreadsheet
• Database Management System (DBMS):
• “Allows user to access and manage the database”
• Examples: Excel, Oracle, MongoDB
• Often also referred to as a database
• Data Warehouse:
• A particular kind of DBMS, “specialized in the data it stores - historic data
from many sources - and the purpose it serves - analytics”
• Azure SQL Data Warehouse, Google BigQuery, Snowflake, Amazon Redshift
Content/quotes from: https://www.dremio.com/what-is-a-data-warehouse/
How Do Data Warehouses Work?

Diagram from: https://www.dremio.com/what-is-a-data-warehouse/

Value of Data Warehouses

Data warehouses provide: Allowing answers to questions like:

• “How much revenue has each product
• “A single point of access for data” line brought in per month over the last
• “An assurance of data quality” 10 years, broken out by city & state?”

• “What is the average transaction size at

• “A history of the data they store”
our ATMs, broken out by time of day &
• “A separation of operations and customer assets?”
analytics” • “What is the % employee turnover for
• “A standard set of semantics the past year in stores that have been
open for at least 3 years. How many
around data”
hours did those employees work per
week?”
All information quoted from: https://www.dremio.com/what-is-a-data-warehouse/
 What is a Data Warehouse?

• Big data tools like Hadoop solve 2 problems:

• Storage of lots of data (usually in a distributed manner) - E.g. HDFS
• Processing big data, which usually involves (i) distributed processing (1000s of
servers), (ii) parallelization - E.g. MapReduce
• Example - when you have text from emails, webpages, etc., and you want to count
the number of times particular words appear
• In this case, Hadoop with MapReduce will:
1. Split the text into groups sent to different machines
2. Map each group into pairs (<key, value>)
3. Shuffle the pairs to group pairs with the same key
4. Reduce each group of pairs with the same key
Content (on 4 steps of Hadoop w/ MapReduce) from: https://hci.stanford.edu/courses/cs448g/a2/files/map_reduce_tutorial.pdf
Images: https://www.diyotta.com/glossary/what-is-hadoop
AI Fundamentals for Non-Data Scientists
Data Analysis: Extracting Intelligence from Big Data

Kartik Hosanagar, Professor of Operations, Information and Decisions

Data Mining

• Term encompassing tools for discovering patterns in large datasets

• Data mining vs. regressions?
• Representation
• Risk = 0.93*prior_default + 0.23*nun_cards – 1.3*employed – 0.734
• How did the technique find the relevant variables and coefficients?
• Regression: analyst had a hypothesis
• Data mining: data-driven exploration
• Data mining techniques/tools
• Clustering 
• Association rule mining
Clustering

• Clustering: Grouping data such that

intra-group similarity is maximized while
inter-group similarity is minimized

• Application: Data-driven customer

segmentation
Association Rule Mining

• Association rule mining - finding common co-occurrences in data

• Basket analysis:
• {bread, butter} ➜ {milk}
• Medical diagnosis: Fish oil and Raynaud’s
disease
• “Local increase of blood viscosity during
cold-induced Raynaud’s phenomenon”
• “Reduction in blood viscosity by
eicosapentaenoic acid”
From Description to Prediction

• Data mining (including clustering and association rule mining) is ultimately

about discovering & describing what patterns exist in data

• The next step, and the core business opportunity, involves using data to
make predictions about the future, called predictive analytics
• Can we predict demand?
• Can we predict fraud?

Content/quotes from: “Prediction Machines” by Ajay Agrawal, Joshua Gans, and Avi Goldfarb
 Predictive Analytics: Amazon Recommendations

Shop and place

Visit website Pay for items Items shipped
in cart

Amazon’s Recommendations

Content/quotes from: “Prediction Machines” by Ajay Agrawal, Joshua Gans, and Avi Goldfarb, Images: https://www.netsville.com/2018/11/13/amazon-hq2-bringing-25k-jobs-to-2-new-east-
coast-locations/amazon-logo-square/, https://icon-library.net/icon/shopping-icon-on-amazon-25.html, https://www.imore.com/amazon-app-everything-you-need-know,
https://services.amazon.com/handmade/getting-started-guide.html, https://www.mageplaza.com/blog/product-recomendation-how-amazon-succeeds-with-it.html
 Predictive Analytics: Amazon Payments

Items Algorithm predicts Payment method

purchased* fraud has occurred rejected

*Past data about customer behavior on the site, coupled with

knowledge of which previous purchases were fraudulent, helps
train the system to make real-time fraud predictions based on
customer behavior (this will be discussed further in Module 3)

Content/quotes from: https://stripe.com/radar/guide, Images: https://visualmodo.com/6-security-tips-protect-ecommerce-site/, https://medium.com/@fenjiro/data-mining-for-banking-loan-

approval-use-case-e7c2bc3ece3
Implications of Big Data & Predictive Analytics

• Big data & predictive analytics have direct benefits for Amazon

• Helpful recommendations ➜ more shoppers and purchases

• Accurate fraud predictions ➜ fewer losses & happier customers

• Perhaps most importantly, prediction can be a virtuous cycle

• Better predictions ➜ more shoppers ➜ more data ➜ better

predictions

Content/quotes from: “Prediction Machines” by Ajay Agrawal, Joshua Gans, and Avi Goldfarb
AI Fundamentals for Non-Data Scientists
Introduction to Artificial Intelligence

Kartik Hosanagar, Professor of Operations, Information and Decisions

Agenda

• Artificial Intelligence (AI) Overview

• What AI is, types of AI, history of AI and the expert systems approach
• Machine Learning (ML) Overview
• How ML differs from AI and the three types of ML (supervised,
unsupervised, and reinforcement learning)
• Detailed View of ML
• ML at 30,000 feet and factors that influence accuracy in ML
• Specific ML Methods: A Deep Dive
• Logistic regression, decision trees & random forests, and neural
networks
What is Artificial Intelligence (AI)?

Definition of AI
• “The theory and development of computer systems
able to perform tasks that normally require human
intelligence” - Merriam Webster
• Understanding language, reasoning, speech
recognition, decision-making, navigating the visual
world, manipulating physical objects, etc.
At its core, AI is about
getting computers to do
things that require human
intelligence.
Images: https://www.securitymagazine.com/articles/90871-whats-the-real-role-of-ai-and-ml-in-cybersecurity
Intelligence on a Spectrum

AI can refer to many different types of intelligence:

Artificial Narrow Intelligence (weak AI):
• AI that is very good at one specific task
• A chess algorithm that can beat any human
• Amazon’s book recommendations
Artificial General Intelligence (strong AI):
• A computer program that could do all intelligent things a human could do, just as
quickly and easily
• Artificial Neural Networks are the closest thing to reaching this level
Artificial Super Intelligence:
• A computer program that can rapidly improve itself and do all things a human
could do at a significant increase in speed and competency
History of AI

Origins
• Alan Turing, Cryptography
• Can machines think?
• “My contention is that machines can be constructed
which will simulate the behavior of the human mind
very closely.”
• Proposed the Turing test for machine intelligence
(1950) - Can machines do well in imitation games?
• Soon after, a workshop called the Dartmouth Summer Research Project on
Artificial Intelligence was held, which historians believe is what coined the
term AI
Content/quotes from: “A Human’s Guide to Machine Intelligence” by Kartik Hosanagar
“ Calling it AI made it extremely ambitious, and it inspired
many people to enter the field, which has been
responsible for a lot of the progress.”
— PEDRO DOMINGOS
AI in the Press

Much of what the general public knows about AI is about AI playing popular
games, which has been widely covered in the press

Images: https://www.nytimes.com/2011/02/17/science/17jeopardy-watson.html
Progression of AI in Games

IBM’s Deep Blue

• May 11, 1997 - IBM’s Chess Playing computer beat #1 in the world, Garry
Kasparov
• No ML in system - Ability to analyze > 200,000 moves per second
IBM’s Watson
• 2011 - IBM Watson beats Ken Jennings and Brad Rutter (two of Jeopardy’s best)
• Using Natural Language Processing (NLP) and Question Answer (QA)
information retrieval
Google + DeepMind’s AlphaGo
• Wins 4:1 vs. World Champion Lee Sedol in GO
• Techniques - Machine Learning (Neural Nets & Reinforcement Learning)
How to Build AI: Expert Systems Approach

• Knowledge engineering, or expert systems

• Capturing and transferring knowledge
from experts to a computer system
• Examples:
• Software to diagnose diseases - interview doctors and codify the rules
they use to diagnose diseases
• Software for self-driving cars - interview drivers and codify the rules they
use to drive cars
• Can create reasonably intelligent systems, but over time expert systems
are often not able to beat human beings at complex tasks that require
intelligence
How to Build AI: Expert Systems Approach (cont.)

Limitations of expert systems

• Polanyi’s Paradox - Tacit knowledge
• Idea that we have a lot of knowledge that we are not aware of
• Asking people to give us all the knowledge they have gets us a good
amount of information, but because of tacit knowledge it doesn’t give us all
the information needed
• Doctors arrive at diagnoses in seconds based off pattern matching, not
rules
• Drivers cannot articulate all the rules they use to drive cars
• The limitations of the expert systems approach has led to the emergence of
a newer approach known as machine learning (ML)
AI Fundamentals for Non-Data Scientists
Machine Learning Overview

Kartik Hosanagar, Professor of Operations, Information and Decisions

What is Machine Learning?

Machine learning (ML) is a subfield of AI

• ML methods are characterized by their ability to learn from data without
being explicitly programmed
• ML is often used for making predictions

PREDICTION TASKS APPLICATIONS OF ML

Structured Data: “Is a transaction fraudulent?” Healthcare: Automated medical diagnosis

Text: “Is an email spam?” HR: Which applicants are best suited for a job
Images: Image recognition in driverless cars Tech: Voice interfaces, Autonomous cars,
personalization
Audio: Speech recognition
Finance: Investing
Three Types of Machine Learning

Supervised Unsupervised Reinforcement

Learning Learning Learning

Acquire new data

Develop predictive Group and interpret
by taking actions
model based on input observations based
and receiving ad
and output data only on input data
hoc feedback

Classification Anomaly Detection Bandit Algorithms

Regression Clustering Q-learning

Supervised Learning

• Idea: Learning from data where you have cleanly labeled outputs and
inputs
• Inputs can be referred to as features/covariates
• Outputs are often called targets (this is what we are trying to predict)
• Example - Email data
• The output we are trying to predict is whether an email is spam or not
• The inputs, or features/covariates, are the actual text in the email
• Supervised learning is used when there is cleanly labeled past data,
which have a “correct” answer
• Certain data have been labeled as “spam”
• Certain other data have been labeled as “not spam”
Supervised Learning Example

Concrete example: Tagging text content on Facebook

• 106,316 Facebook posts submitted by large brands; wanted to identify what types of
posts were associated with highest engagement
• Manually tagging content is expensive, so we wanted to automate it with a supervised
machine learning technique
• Solution - have a random sample tagged manually on Amazon Mechanical Turk, and
use as a training dataset for supervised ML
• Took a sample of 5,000 posts and had human beings label each of these posts
• These labels were then used as a training data set for a supervised machine
learning algorithm that learned what kinds of words are predictive of whether a post
is emotional, or humorous, etc.
• That algorithm was used to make predictions for the remaining posts that hadn’t
been labeled by a human being
NLP Algorithm Performance

Algorithm described in detail at: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2290802

Unsupervised Learning

• No fixed set of outputs provided; must be learned from inputs

• Goal is to cluster and identify important features
Unsupervised Learning

• Latent Dirichlet Allocation (LDA)

• Takes as input a set of documents
• Identifies both common topics across set of documents
• Labels documents with corresponding topics
Unsupervised Learning
Reinforcement Learning

• Idea: Let algorithms learn by testing various actions/strategies and deciding

which one works best
• Very powerful method for simulation and robotics-based applications
• At the heart of AlphaGo and other state-of-the-art gaming AI
• In many applications experimentation isn’t free
• You often want to balance both exploration and exploitation
• Multi-armed bandit algorithms
Reinforcement Learning

• Suppose you have two ad copies and you don’t know which will attract
more clicks (and therefore visitors to your website)
• Traditional A/B testing involves showing ad A 50% of the time and ad B 50%
of the time, and then assessing which ad performed better

Ad A Ad B

B performs better than A

The A/B test line is midway between the A
and B lines as it is 50% A and 50% B
Reinforcement Learning

• Machine learning can improve upon A/B testing through bandit algorithms
• Bandit algorithms update beliefs based upon performance
• They spend more time on best
performers early on while still
learning and improving over time
• The bandit begins by showing
50% A & 50% B, but slowly starts
allocating more and more traffic
to the higher-performing ad as it
learns and confirms which one is
better
AI Fundamentals for Non-Data Scientists
Reinforcement Learning

Kartik Hosanagar, Professor of Operations, Information and Decisions

Bandit Algorithms for Website Optimization

Website “Action” space Objective

Conversion rate

Revenue per session

E-commerce retailer Design of product page Outcome of interest

Exploration and Exploitation

• Exploration - gathering more information about the decision environment

• Exploitation - making the best decision given the current information
• Example: Restaurant selection
• Exploration - going to a new restaurant
• Exploitation - going to your favorite restaurant
• How do we balance exploration vs. exploitation?
• This tradeoff is handled by algorithms like multi-armed bandit, a classical
reinforcement learning approach

Content/quotes from: https://www.davidsilver.uk/wp-content/uploads/2020/03/intro_RL.pdf

Multi-Armed Bandit

A finite set of resources must be allocated among multiple choices

• Gambler has finite amount of time

• Try new slot machines to try to maximize reward (exploration)
• Stick with the slot machine producing reasonable returns (exploitation)
Content/quotes from: https://www.amazon.com/RecZone-Jumbo-Slot-Machine-Bank/dp/B000H0KBDY
Bandit Algorithms

Many algorithms can be used to balance exploration and exploitation

• Epsilon First - Heuristic in which we experiment (explore) early,

once we’ve learned, start exploiting
• Allocate 100% of traffic to the best performing variant

• Thompson Sampling - Allocate traffic probabilistically to

each arm, with higher probabilities given to arms that have a higher chance
of being the best
Summary

• Reinforcement learning offers an alternative to ML based on having large

datasets
• Reinforcement learning has found many applications in gaming and in
online personalization
• Reinforcement learning is not as widely used as supervised machine
learning
AI Fundamentals for Non-Data Scientists
A Detailed View of Machine Learning

Kartik Hosanagar, Professor of Operations, Information and Decisions

ML at 30,000 Feet

Focus on supervised learning:

• At its core, supervised machine learning is about using “a set of variables” to
predict “an outcome”
• You do this all the time
• Dark clouds + strong winds ➜ rain
• What someone is wearing, how they interact with you ➜ whether you’ll be
friends
• Resume (school, experience, skills) ➜ good employee
• Applications in business are abundant
• Will someone buy your product?
• Will someone click your ad?
ML at 30,000 Feet

Focus on supervised learning:

• At its core, supervised machine learning is about using “a set of variables”
to predict “an outcome”

UserID PgsViewed ZipCode OS Purchase

1 3 19104 iOS Yes
2 10 19141 Windows No
… … … … …

• Names for input data - predictors, features, “data”, variables,

characteristics, covariates
ML at 30,000 Feet

Focus on supervised learning:

• At its core, supervised machine learning is about using “a set of variables”
to predict “an outcome”

Var1 Var2 Var3 Var4 Outcome

ML at 30,000 Feet

Focus on supervised learning:

• At its core, supervised machine learning is about using “a set of variables”
to predict “an outcome”

Var1 Var2 Var3 Var4 Outcome

ML at 30,000 Feet

Focus on supervised learning:

• At its core, supervised machine learning is about using “a set of variables”
to predict “an outcome”

Var1 Var2 Var3 Var4 Outcome

All of supervised ML comes down to

approximating this function with high fidelity
What Influences Accuracy in ML?

What is accuracy?
• Ability to make correct predictions on data you haven’t seen

Var1 Var2 Var3 Var4 Outcome

What Influences Accuracy in ML?

Quantity of data
• Number of distinct observations

Var1 Var2 Var3 Var4 Outcome

What Influences Accuracy in ML?

Quantity of data
• Number of distinct observations

Var1 Var2 Var3 Var4 Outcome

What Influences Accuracy in ML?

Quantity of data
• Number of distinct observations
• Different characteristics about your observations

Var1 Var2 Var3 Var4 Outcome

What Influences Accuracy in ML?

Quantity of data
• Number of distinct observations
• Different characteristics about your observations

Var1 Var2 Var3 Var4 Var5 Var6 Var7 Var8 Outcome

What Influences Accuracy in ML?

• Quantity of data (number of rows)

• Number of features (columns)
Lots of other things!
• Relevance of information to underlying phenomena
• Umbrellas predict rain better than color of a person’s dress
• Complexity of your model (how we approximate “f”)
• Modern developments in "deep learning" allow for very complex models
that were historically very hard to train
• “Feature engineering”
• Using domain knowledge to create new features from the input data
(some of it can be automated but the analyst has a big role to play)