Artificial Intelligence Data science introduction

Data

1. The Landscape of AI
Shankar Venkatagiri
Indian Institute of Management Bangalore

2. Policies
•Quizzes: All quizzes & exams on DigiExam, no makeups
•Closed book, closed notes, closed Internet
•Let’s keep this class interactive
•No side conversations
•Laptops & mobiles shut unless required
•Bio & water breaks before class start
•Attendance may be audited at random
•TA: Rahul Pon, Doctoral Candidate, IS Area
•GPU Chief in Residence, Demo creator, …
Midterm Final Quizzes Project
35 35 10 20

3. EXISTENTIAL
➤Video: Jon Stewart on AI

4. ➤Q: How to identify fraud?
PREVIOUSLY

5. ➤Q: How to identify fraud?
➤Option: Brainstorm with “experts” ➠ prescribe rules
➤Run each transaction by these rules, evaluate
Rule 1: Amount > 3 * 365-day average? Fraud
Rule 2: Location > 500 miles from Residence? Fraud
…
…
Rule 293: Date of Last Usage > 1.5 years? Fraud
PREVIOUSLY

6. ➤Q: How to identify fraud?
➤Option: Brainstorm with “experts” ➠ prescribe rules
➤Run each transaction by these rules, evaluate
Transactions Apply Rules Evaluate
Launch!
Handle errors
!
"
PREVIOUSLY

7. THE ML WAY
➤Train a model with historical data
➤Learn the rules, apply them, evaluate outcomes, and deploy the model…
Data
Transactions Evaluate
Launch!
Handle errors
!
"
Train ModelApply Rules
Machine
Learning
Rules
Answers
Data
E.g. Classifiers

8. ➤Model coefficients can be calculated through formulas
PREVIOUSLY
Model: y = 0.27 + 0.7 x
^
Data
❖Estimated y = b0 + b1 x
Slope b1 = SSXY / SSX
Intercept b0 = y - b1 x
SSX = ∑xi
2 - (∑xi)
2 /n ; SSY = ∑yi
2 - (∑yi)
2 / n
SSXY = ∑xi yi - (∑xi)(∑yi)
/ n
Rules
Answers

9. THE AI WAY
➤Learn from data and answers, discover the rules
➤Start with the data ⇒ Propose a random model ⇒ Arrive at “best” coefficients
➤Each iteration reduces a “loss function” - e.g. MSE

10. ANALYTICS
➤Analytics: deriving insights from data and making business decisions
➤Descriptive (Simulation) / Predictive (Forecasting) / Prescriptive (LP)
➤All approaches rely on an efficient pipeline
➤Managers deliberate about building models, deploying them, and monitoring
Acquire Data Clean Pre-process Feature Engineering
ModelEvaluateDeploy
Monitor Model
and Update

11. ➤Supervised vs Unsupervised - labels available?
➤Reinforcement learning - goal, rewards & penalties
MODELS
Tasks Approach Models
Estimating the price of a house
Predicting the next ten stock movements
Regression Regression, RNNs
Identify spam mails
Detecting a fraudulent transaction
Classification
Decision Trees,
SVM, Naive Bayes
Finding products often bought together Association Mining Apriori
Identify customer groups with similar buying patternsClustering K-Means

12. STACK
➤AI - Intelligence demonstrated by machines
➤Goal: Automate intellectual tasks performed by humans
➤General purpose technology, problems tackled are complex
➤E.g. Tesla auto-drive, Alexa translate, …
➤Neural networks try to mimic the brain
➤Q: How many neurons in a human brain?
Artificial
Intelligence
Machine
Learning
Deep Learning

13. PROBLEM
➤Q: How does this appear to a machine?
➤Challenge: Given this 28 x 28 matrix,
match it with a digit between 0 and 9
Answer
= 4

14. SMART
➤Q: Is this hype, or are folks in Bengaluru using AI?
➤Video: AI to assist Police with Bengaluru traffic (upto 8:45)

15. DEEP
➤Deep Neural Network (DNN) - artificial neurons connected in layers
➤Tasks: Classifying images, generating music, …
➤Learning involves adjusting layers & edge weights to achieve the task
➤Flavours: CNN, RNN, GAN…
Inputs Outputs
Hidden layers
of neutrons
Source: Chollet, F. Deep Learning in Python (2018)

16. CHATGPT

17. GEMINI
➤Q: How much do we trust AI?