1 of 24

Lecture 1:

Theory

Lecture 2: Code Walk-Through

Hands-on Session: Small Group

Fundamentals in in Neural Networks (Regression)

Linear Regression, Training/Testing

Hands-on work

Hands-on work

Hands-on work

Hands-on work

Hands-on work

Logistic Regression

Polynomial Regression, Overfitting, and Tuning

Week 1

Week 2

Week 3

Week 4

Week 5

Intro to Data Science & Exploratory Data Analysis (EDA)

Polynomial Regression, Tuning a Model

EDA, Train/Test Split, Linear Regression

Logistic Regression & Multiple Logistic Regression

Introduction of Tensorflow Keras and Neural Networks

Intro to Python & Basic Programming

Course Structure

2 of 24

Closing

Ceremony

Project: Baseline Model

Presentations

Project work

Project: Advanced Model (Upgrade from Baseline)

Week 6

Week 7

Week 8

Week 9

Week 10

Convolutional Neural Networks (CNNs)

Tools for Improving CNNs: Regularization and Transfer Learning

Image Classification with CNNs

Ethics in AI

Advanced Topics in Image Classification: Using VGG16

Project work

Project work

Tuning Neural Networks (Classification)

Project: Start Projects with EDA!

Tuning NNs, Using NNs for classification, Validation Sets

Course Structure

3 of 24

CAPSTONE PROJECT

  • Group project with teams of 3 - 5

  • Covers a real-world topic in AI, using a real-world dataset

  • You get to submit project choice preferences

  • Each team presents on their project during final week

4 of 24

Chat Waterfall: What is AI?

5 of 24

What is AI?

Vision

Composition

Drug discovery

Diagnosis

6 of 24

What is AI?

Vision

Composition

Drug discovery

Diagnosis

  • AI finds patterns in complex data that often we cannot find on our own!
  • Then, uses those patterns to make predictions or classifications.

7 of 24

To make the best AI, you first need to do good data science. Ask questions like…

  • Where did this data come from?
  • Is my sample generalizable?
  • Is this relationship correlation or causation?
  • How can I interpret my model?

8 of 24

Let’s see an example of how knowing your data leads to better predictions…

9 of 24

Abraham Wald and the Allied Forces Statistical Research Group

10 of 24

Image courtesy of Professor Joseph Blitzstein (i.e. the best stats prof in the country). See his “Harvard Thinks Big” talk on this problem: https://youtu.be/dzFf3r1yph8

Chat Waterfall: Where would you put extra armor?

a) Nose (front)

b) Main wings

c) Main body

d) Engines

e) Tail (back)

11 of 24

Image courtesy of Professor Joseph Blitzstein (i.e. the best stats prof in the country). See his “Harvard Thinks Big” talk on this problem: https://youtu.be/dzFf3r1yph8

Where would you put extra armor?

a) Nose (front)

b) Main wings

c) Main body

d) Engines

e) Tail (back)

Wald’s Answer

12 of 24

Image courtesy of Professor Joseph Blitzstein (i.e. the best stats prof in the country). See his “Harvard Thinks Big” talk on this problem: https://youtu.be/dzFf3r1yph8

These are the planes that came back.

So these dots are places where planes can take hits and still make it back.

13 of 24

Image courtesy of Professor Joseph Blitzstein (i.e. the best stats prof in the country). See his “Harvard Thinks Big” talk on this problem: https://youtu.be/dzFf3r1yph8

The planes not included in our sample – the ones that went down – may have looked like this.

14 of 24

The Plane Problem:­

Sample Bias

15 of 24

Where does sample bias come up in modern AI problems?

16 of 24

Models built from existing stock data ignore firms that left the market (& the factors that made them go broke).

Predicting Stock Prices

Self-Driving Cars

Perform worse on roads that look very different than the ones they’ve seen before

Facial Recognition

Shows bias if trained on datasets with predominantly white faces

17 of 24

Exploratory Data Analysis (EDA)

18 of 24

EDA is critical to all data projects (and will be needed for your projects in a few weeks!)

EDA helps us:

  • Understand the variables in our data
  • Select the modeling tools
  • Can even help us generate hypotheses!

Why Exploratory Data Analysis (EDA)?

19 of 24

Categorical Variables

  • Categories that don’t have an inherent order
  • Doesn’t make sense to “average” them (e.g. you can’t find the average of several clothing types)
  • Often expressed in terms of counts

Rapid Fire Chat: List some examples!

20 of 24

Quantitative Variables

  • Quantities that can be ordered
  • Makes sense to “average” them (e.g. the average height in the room)
  • Can be continuous (weight of lumber) or discrete (number of cars people own)

Rapid Fire Chat: List some examples!

21 of 24

Pop Quiz

Chat Waterfall: Are zip codes categorical or quantitative data? Why?

22 of 24

EDA for Categorical Data

One Variable Visual

Barplot

Two Variable Analysis Two-Way Tables + Barplots

Summary Stat Proportions

Percents of sample:

  • Blue - 22%
  • Red - 36%
  • Orange - 30%
  • Pink - 12%

23 of 24

EDA for Quantitative Data

One Variable Visual

Histogram or Boxplot

Two Variable Analysis Scatterplot + Correlation

Summary Stat 5-number summary

Heights of people

  • Min: 48in
  • Q1: 56in
  • Med: 64in
  • Q3: 68in
  • Max: 74in

r = 0.76

24 of 24

Correlation