1 of 25

Week #1 : Introduction

Human - AI

Interaction

Human - AI

Interaction

Human - AI

Interaction

​

​

Chinmay Kulkarni and Mary Beth Kery

Fall 2019, Human-Computer Interaction Institute, Carnegie Mellon University

Human-AI Interaction Fall 19 .

2 of 25

Who we are

Chinmay Kulkarni, Assistant Prof HCI

Studies and builds systems for large-scale learning, work, and mentoring.

Mary Beth Kery, 5th year PhD student HCI

Studies how people code with AI/ML and builds developer tools for experimentation.

2

Human-AI Interaction Fall 19 .

3 of 25

Why we’re doing this course

  • Most AI/ML courses consider “user-interfaces” or human impact as an afterthought; and focus narrowly on algorithms

3

Human-AI Interaction Fall 19 .

4 of 25

Chinmay’s personal experience with learning AI

Sep 2010: Third week of class.

​

​

​

​

​

I’ve used this beautiful mathematical result zero times for building interactive AI

4

Human-AI Interaction Fall 19 .

5 of 25

Mary Beth’s personal experience with learning AI

5

https://experiments.withgoogle.com/collection/ai

glitch art built on pix2pix + tensorflow

Human-AI Interaction Fall 19 .

6 of 25

On the other hand, algorithms are not always the answer

​

  • Recommender systems: If you go to Netflix for the first time, what should it recommend you watch? [The cold start problem]
  • Mathematically, no one has solved the cold start problem with any meaningful guarantee. User experience design is critical to help overcome limitations.
  • Cold start Quora: Cold start iOS face id:

​

6

Human-AI Interaction Fall 19 .

7 of 25

Why we’re doing this course

  • Most AI/ML courses consider “user-interfaces” or human impact as an afterthought; and focus narrowly on algorithms
  • My experience at a startup: The only way to win is to cheat

7

Human-AI Interaction Fall 19 .

Human-AI Interaction Fall 19 .

8 of 25

Chinmay’s personal experience with AI startups

  • I have advised a few.
  • Most start with huge aspirations:
    • We will completely automate [PR, sales, customer service, driving…]
    • The technology seems to be there?
  • But too often, challenges arise:
    • Don’t figure out how to interact with people
    • Aren’t able to collect data
    • Etc.
  • The only way to win is to “cheat”: don’t think of the AI that can solve it all, but the human+AI team that will.

​

8

Human-AI Interaction Fall 19 .

9 of 25

Why we’re doing this course

  • Most AI/ML courses consider “user-interfaces” or human impact as an afterthought; and focus narrowly on algorithms
  • My experience at a startup: The only way to win is to cheat
  • With you, we are co-inventing a new user-centric way to build AI systems

9

Human-AI Interaction Fall 19 .

10 of 25

AI affects many facets of human life & society

10

Human-AI Interaction Fall 19 .

11 of 25

Earlier this year… (by the person leading Skype)

11

Good: train your model is tiny

Bad: where are all the humans???

Human-AI Interaction Fall 19 .

12 of 25

Why we’re doing this course

  • Most AI/ML courses consider “user-interfaces” or human impact as an afterthought; and focus narrowly on algorithms
  • My experience advising startups: The only way to win is to cheat
  • With you, we are co-inventing a new user-centric way to build AI systems - let’s do it!

12

Human-AI Interaction Fall 19 .

Human-AI Interaction Fall 19 .

13 of 25

Our teaching philosophy

  • Put people first!
  • AI/interaction NOT from the basics
    • Instead, we will use building blocks which are easier to reason about. Don’t worry -- these building blocks are state of the art!
    • When we talk about algorithms, we will anthropomorphize: e.g. linear regression just wants to draw a line in the sand. Logistic regressions wants to pick a zero or one, and is unhappy in the middle. Given a lot of choices, Softmax just wants to pick one.
      • It’s weird but effective.
  • We’ll help you think it through, but also “do it through” development
    • Nothing gives you a sense of possibilities and limitations like building it
  • To learn something, teach it
    • True in general, but for AI, industry folks have all sorts of ideas (often wrong).
    • To succeed, you need to teach them

13

Human-AI Interaction Fall 19 .

14 of 25

What should you expect to do in this course?

  • Thinking it through: Readings, reflections, quizzes
  • Doing it through: Assignments and projects
  • Teaching it through: In class panel discussions, debate, etc.

14

Human-AI Interaction Fall 19 .

15 of 25

Doing it through: projects

  • Assignment 1: build a system to determine if someone should get a mortgage, see how well it works, and for whom
  • Assignment 2: Build UI around mortgage system
  • Assignment 3: Visualize data that goes into an AI system. See why it matters
  • Assignment 4: Build a chatbot to recommend stuff to you
  • Assignment 5: User-facing computer vision application
  • Final project: Make something cool and interactive with AI/ML (or write about how people use a particular AI/ML product)

15

Human-AI Interaction Fall 19 .

16 of 25

Do you know enough programming?

You need to know some Javascript and Python

Example tasks:

Javascript: Click a button, show a modal

Python: Take data as a csv, and transpose all rows into columns

​

16

Human-AI Interaction Fall 19 .

17 of 25

What questions do you have?

17

Human-AI Interaction Fall 19 .

18 of 25

What do we mean by AI?

Computers doing things that we expect people to be able to do

  • Rule-based AI/expert systems: mimic an expert.
  • Planning/Solving: e.g. compute directions from CMU to IKEA.

Learning from examples: machine learning, computer vision, natural language processing etc.

18

Screenshot from http://www.aaai.org/Papers/AAAI/1982/AAAI82-070.pdf

Human-AI Interaction Fall 19 .

19 of 25

Make it through

Create a journey map from getting lunch to coming to class. Where did AI come in, where did it not?

​

Where should it have? Where should it not?

19

Human-AI Interaction Fall 19 .

20 of 25

20

Human-AI Interaction Fall 19 .

21 of 25

Journey map: What we’ll do now (Part 1)

  1. (7 minutes) Make your own personal journey map -- alone
  2. (3 minutes) Label on your map:
    1. Where did AI come in?
    2. Where should it have?
    3. Where should it not have
  3. (5 minutes) Discuss with 3-4 neighbors
    • What do you see as common themes?
    • How did your labels differ from others?
  4. (5-8 minutes) Share with the class:
    • common themes
    • Major points of difference

21

Human-AI Interaction Fall 19 .

22 of 25

Analysis (Part 2)

  • (4 minutes) Discuss: The common themes where AI does or should come in:
    • how might it fail?
    • What concerns do you have?
  • (4 minutes) Alone: How might you fix one of these concerns
    • Can you create a purely technical solution?
    • Can you create a solution part-technical, part people?
  • (5 minutes) Discuss: Solutions
    • What do you see as common themes among solutions?
  • (5-8 minutes) Share with the class:
    • common themes
    • Unsolved challenges

22

Human-AI Interaction Fall 19 .

23 of 25

Affinity diagram (Part 3)

  1. (10 minutes) Come up to the whiteboard, put up post it notes for your solution:
    1. Challenges: Ethics, privacy, transparency, oversight, etc.
    2. Technical issues: How to present to users, how to collect data, how to evaluate, how to handle failures, etc.
    3. Applications: Recommender systems, natural language processing/understanding, computer vision, humans-in-the-loop, etc.

​

23

Human-AI Interaction Fall 19 .

24 of 25

Course trajectory

  1. Week 1-2: Orientation, and historical perspectives
  2. Week 3-6: How to interact with users, how to collect data
  3. Week 5-9: The big societal questions: ethics, data privacy, and more.
  4. Week 10-15: Specific application areas: Recommendations, NLP, Vision, etc.

​

24

Human-AI Interaction Fall 19 .

25 of 25

One more thing

  • Office hours for Chinmay: Prefer Monday or Friday?
  • What other questions do you have?

25

Human-AI Interaction Fall 19 .