CS 470 - Spring 2020.

Welcome to CS 470!

Lecture 2: Introductory Lecture. Part II. 1/15/2020

Administrivia

I have office hours Wednesdays from 4-6 pm, Room 014 AKW. Also, I invite students to join me for lunch in a residential college after class on Wednesdays. Today at Franklin at 11:45am.

The course assumes proficiency in python programming. We will spend a few lectures on python, but generally expect you to accomplish that on your own, if you are not already an adept. If you would like some help, as well as course credit, in learning python, we recommend CS 200, which meets MW 2:30pm in WLH 120. We hear that the instructor is quite charming. You can drop by this afternoon.

If you're thinking of taking CPSC 470, please do the following.

Complete this online Student Information form.

Get a course account for CPSC 470 on the Zoo. When you register for this class, your course account should be created within one hour or so of signing up.

Confirm your piazza account. Yesterday, I created piazza accounts for the class. You should have received an email requesting that you validate your account. The guys at Canvas have alerted the Yale community that Piazza now automatically shares your information with potential employers and their ilk. ( how to turn this feature off.)

Start the homework assignment: [Assignments]. It is a python refresher.

Lecture: Introduction to AI

AI has a yin yang duality. Most modern AI programs comprise focussed applications to specific domains, or collections of such focused applications. They are often based on quantitative, statistical techniques. Terms such as big data, machine learning, or deep learning generally refer to this flavor of AI.

By contrast, classic AI, or what we will, with minimum vanity, call classic Yale AI, tries to replicate human cognitive behavior with the twin goals of creating smart programs and also gaining an understanding of how humans think.

One of the early tasks for AI was chess playing. Smart people play chess. If a computer could play chess and beat a smart person, the computer would be intelligent. Right? Simon and Newell, early AI researchers at Carnegie Mellon, tried creating computer programs based on verbal protocols of expert chess players. That is, the programs tried to replicate the stated mental processes of people. These programs were OK, but not great. The dominant approach to game playing (which we will explore in this course) involves searching the game tree to the greatest depth possible. This algorithm is not what humans do, but it proved successful in defeating humans. Search is a cornerstone of modern AI, and computer science in general. Think Google.

Most of the techniques in the textbook and course are of the latter variety. How can we get a computer to accomplish a task, even if there is no evidence that people use that method?

Modern AI

2019 Intro Slides These slides are from Professor Scassellati. (Slide 30)

Ford home delivery robot What could go wrong?

Hitchhiking robot

Classic Yale AI

The second approach is how can we get computers to replicate human cognitive behavior. That is, we assume that human cognition comprises a vast number of processes and algorithms. What are those processes and can we embody those in a computer?

The goal is not only to create smart programs, but also to provide insights into how people think. This is a cognitive science approach, which has a variety of benefits.

A computer that thinks like a person should be better at communicating with a person.
A computer that makes decisions as a person would should be able to explain the decision to a person and perhaps persuade a person to change her mind.
A computer that learns as a person does could provide the basis for programs that are better at teaching and educating people.
A computer that understands human interpersonal relationships might be able to use that as a model for relating to other intelligent agents - both carbon and silicon based.
A computer that understood and even displayed emotions should be better at communicating with people, pace Mr. Spock.

Classic Yale AI began in 1974 with the arrival Roger Schank in computer science, who collaborated with Robert Abelson in psychology. Here are the players.

Roger Schank. Created conceptual dependency, a language of thought independent of any natural language.
Robert Abelson. Developed script theory with Schank.
Robert Wilensky. Student of Schank and Abelson who developed theory of plans, including the persuade package. After Yale, he was a professor at Berkeley.
James Meehan. Student of Schank who wrote story generation program TALE-SPIN based on planning theory. See examples and mis-spun tales and grand text auto. His thesis was entitled The Meta Novel.
Drew McDermott. Moved to Yale from MIT in 1976. He taught this course for many years. See Artificial Intelligence meets Natural Stupidity Drew McDermott, 1976.
Ernest Davis. Student of McDermott. On the faculty at NYU. See AI Is Harder Than You Think New York Times, 2018.
Stephen Slade. Student of Schank, Abelson, and McDermott. Created VOTE, an interpersonal, goal-based model of decision making.
Peter Norvig. Student of Wilensky (at Berkeley). He is Director of Research at Google and the co-author of our textbook.
There are dozens of others, but these are the usual suspects.

You are familiar with the Turing Test for AI, which is often dismissed these days. The Turing Test was largely replaced in the academic community by Searle's Chinese Room, which was inspired by a 1979 visit to the Yale Artificial Intelligence Lab.

Original paper: Minds, Brains, and Programs Behavioral and Brain Sciences, 1980. Note especially responses by Abelson, Schank, and Wilensky.
Consciousness in Artificial Intelligence video of John Searle talk at Google, 2015. At around 9 minutes, Searle talks about visit to Yale AI Lab.

Modern AI and Classic Yale AI are two perspectives. You may view them in Hegelian terms and thesis and antithesis. In this course, we argue for the synthesis. Other dimensions of the duality include

orthodox vs iconoclastic.
respectful vs irreverent.
quantitative vs qualitative.
neat vs scruffy. (Abelson's terms. It is of interest that Abelson was a world class statistician. Yet, he was attracted to qualitative.)

CS 470 Thought Experiment Please complete. The purpose of this is to get you thinking about the topics, not to find the correct answer.

Next steps: getting up to speed in Python and jupyter notebooks.

[Home]