From 6.034 Wiki

Revision as of 19:50, 17 August 2020

Welcome to the 2019 Edition of 6.844

Overview

6.844 was created in response to requests from grad students who wanted to take 6.034, but needed graduate level credit.

It is a supplement to 6.034---you will take 6.034 as usual and do all of that work (lectures, labs, quizzes), and in addition attend the 6.844 session and do the work required there. That session will meet every Friday 11am-12pm. Each week there will be a reading assignment focusing on one or more of the foundational, provocative, or intriguing papers from the research literature. You will be expected to do the reading, write up a one page response to a set of questions that will be provided with the reading, and come to class prepared to discuss your (and others') answers to those questions.

The papers will help you learn how to read original research papers in the field and will focus on the science side of AI, addressing the larger scientific questions, rather than existing tools for building applications.

The class is heavy on interaction; you will not be able to just sit back and listen. To keep the class size manageable and to encourage active class participation, we do not allow listeners.

More information about the class can be found here.

Staff

Prof. Randall Davis Instructor davis@mit.edu	Jack Cook Teaching Assistant cookj@mit.edu

Week 1 -- September 13

The paper below is for discussion on Friday, 13 September:

"Steps Toward AI" by Marvin Minsky, available here.

A few comments to guide your reading:

Keep in mind first that this paper was written in 1961, 58 (fifty eight!) years ago.

As the guest editor’s comment indicates, this is very early in the birth of the modern version of the field; Minsky had been invited to write a tutorial overview.

Recall that your job is to summarize the paper in one page. Do that, and also try to comment on these things as well:

1. How many of the ideas Minsky mentions do you recognize as still in use?

2. Does he do a good job of laying out the structure of the field?

3. What is that structure?

4. Consider the sentence near the top of page 9 beginning “A computer can do, in a sense.…” There are several reasons why he starts off that way. List some reasons that seem compelling to you.

Week 2 -- September 20

Note: There is no class on Friday, 20 September. It's a student holiday.

The paper below is for discussion on Friday, 27 September. We will discuss the evolution of rule-based expert systems. The discussion will be based on your comments and insights on this paper:

Robert K. Lindsay, Bruce G. Buchanan, Edward A. Feigenbaum, and Joshua Lederberg. "DENDRAL: A Case Study of the First Expert System for Scientific Hypothesis Formation" in Artificial Intelligence 61, 2 (1993): 209-261.

The paper is available here.

This is a rather long paper, but you have extra time to review it and to consider the interesting ideas in it. Given the size and depth of the paper, it would be a bad idea to wait until the last minute to read it.

A reminder from the overview info about the course:

Your weekly write-up should not be longer than one page and should be readable by someone who hasn't read the paper. The ability to write such a review is an important skill to develop. The idea is not to include a pile of mathematical formulas or lots of code in your review. We want you to learn to extract the essential take-away message of the paper, including such things as:

1. What is the author trying to accomplish, i.e., what is the problem they are trying to solve? Why is it difficult?

2. What technical methods is the author bringing to bear?

3. Did the work succeed? What does “succeed” mean in this case?

4. If it worked, why did it work? Where it failed, why did it fail? (Failures are typically among the most interesting and revealing behaviors of a program.)

Week 3 -- October 4

This paper is for discussion on Friday, 4 October: "What Are Intelligence? And Why?" by Randall Davis, available here.

This is another overview of AI paper, rather than a technical examination of a particular technique or program. It tries to take a step back and answer a core question: What is it that we're talking about when we talk about intelligence? The paper suggests that intelligence is many things and has been interpreted differently from several different intellectual foundations.

Your task is to evaluate how successful the paper is in answering the questions it raises. And pay no attention to the name of the author. I expect you to be hard-headed and clear-headed in your evaluation and/or criticism.

Week 4 -- October 11

The paper below is for discussion on Friday, 11 October.

You've recently been learning about deep neural nets, which have been strikingly successful in computer vision, speech understanding, and a range of other classification tasks. But there is also an interesting problem with them. Deep Neural Nets are Easily Fooled, which is capitalized because that's the title of a very interesting paper, available here.

If you are off-campus, that link might not work, in which case use this one.

You may also wish to look at this web page.

Keep in mind though that I've seen it as well, so just repeating what you see there will not be considered a good write up.

As usual the point is not just to simply summarize the paper, but to think about what interesting ideas are in there, describe those, and then evaluate them. Make your own judgments, and tell me what you think and why.

Week 5 -- October 18

Given the other things going on this week, I have selected a more non-technical paper to read. It's still challenging and needs some thought, but it's also fun. It concerns a famous argument about the possibility of computers thinking. Read it over and explain how you react to the arguments. As usual, don't just summarize the paper; summarize the issues but then describe your response to them. Are they convincing? If so, why? If not, why not?

The paper is here.

Note that there is no easy or obvious answer here, the idea is to take the argument seriously and think about how you might respond.

Week 6 -- October 25

This week's reading seems a nice way to continue our discussion of what we mean by "understand", whether talking about people or about programs. It appeared online the day before our last class and mentions the Chinese Room thought experiment (which is why it came to the attention of one of our class members, who sent me a pointer).

It discusses the efforts to design successively more challenging language understanding tasks, as a way of working toward the goal of programs that understand natural language. And hence more generally, understand.

One piece of background you'll need is a basic understanding of word embedding. This brief article summarizes a research paper (cited at the end). It introduces the idea, shows you one of the very well known examples, and explores how robust this technique is:

[1]

In a throwback to primary school, you'll also need to understand sentence diagramming. A quick intro (or refresher):

[2]

Now imagine that a sentence is just a noun phrase followed by a verb phrase. In your writeup, write this sentence with parentheses that separate it into the noun phrase and the verb phrase:

The tall person who owns the hammer that is too heavy for Sam to lift likes chocolate.

Notice how much simpler the sentence is to understand once it's divided up into its constituent parts.

Notice how two words -- "person" and "likes" -- that are very far apart in the original sentence can now far more easily be seen as connected to one another.

You'll need this to understand the part of the article that describes treelike representations of sentences and neural nets trained with non-sequential representations.

With that background, the main article for this week is from Quanta Magazine:

[3]

Points for your writeup (in addition to the sentence re-writing above):

1. Explain the challenge tasks the paper describes in the efforts to reach a better level of understanding in software. Do these seem to be working? Why or why not?

2. These efforts are sometimes described as a cat and mouse game. How does that apply here?

3. Evaluate the claim the neural network building is now a well defined engineering practice, in the sense that the right architecture is easily determined, built and trained. If not, why not?

4. Consider this quote from the article:

The only problem is that perfect rulebooks don't exist, because natural language is far too complex and haphazard to be reduced to a rigid set of specifications. Take syntax, for example: the rules (and rules of thumb) that define how words group into meaningful sentences. The phrase "colorless green ideas sleep furiously" has perfect syntax, but any natural speaker knows it's nonsense. What prewritten rulebook could capture this "unwritten" fact about natural language -- or innumerable others?

Presumably you understood the "colorless green..." sentence as meaningless in the literal sense (i.e., leave aside poetic interpretations). How did you do that? That is, how did you do that in a way that would allow you to do it for innumerable other such sentences? Do you have a rule book full of unwritten facts in your head? If not, how did you figure out that this sentence (and others like it) are problematic? What do you know?

Week 7 -- November 1

You've been learning about near miss as an interesting model of learning. One of the issues in this technique is the origin of the near misses -- where do they come from? How do we know what near misses to supply?

This paper, mentioned briefly in the 6.034 lecture, offers a real use of near-miss learning to inform a sketch recognition system that works from descriptions of shapes. The difficult part is getting those descriptions and getting them to be correct.

https://ai6034.mit.edu/wiki/images/HammondDavisIUI06.pdf

We've read aseveral overview papers thus far in the course; this one is a description of a particular system and a particular approach to learning. The framework that I supplied at the beginning of the term even more relevant in this context:

1. What is the author trying to accomplish?

2. What technical methods is the author bringing to bear?

3. Did the work succeed? What does “succeed” mean in this case?

4. If it worked, why did it work? Where it failed, why did it fail? (Failures are typically among the most interesting and revealing behaviors of a program.)

And remember, don't read the paper back to me; read it and think about it and evaluate the ideas.

Week 8 -- November 8

In last Monday's lecture I gave an overview a core topic in AI -- representation -- that has been somewhat overshadowed recently as a result of the interest in statistical models (e.g., neural nets, SVMs, etc.) It's important to understand how knowledge might be represented explicitly in a program, rather than indirectly (e.g., in the weights of several million neurons). The 6.034 lecture reviewed several of the key representations that have been developed.

The paper for this Friday is an overview of the topic of representation in general, trying to consider what a representation, is by thinking of the fundamental roles it plays.

The paper lays out 5 such roles -- for each of them:

1) explain what the role is

2) explain how it does (or does not) help make clear to you the concept of knowledge representation. Well reasoned critiques are welcome here. Is there another role that should have been used to explore the concept?

https://ai6034.mit.edu/wiki/images/Whatkr.pdf

(You will find Table I of the paper familiar; it was used in the later paper "What Are Intelligence" that we read earlier in the term.)

Week 9 -- November 15

In 6.034 we have seen several models of learning (nearest neighbor, neural nets, SVMs, etc.). All of them share the property that they require large numbers of examples in order to be effective. Yet human learning seems nothing like that. It's remarkable how much we manage to learn from just a few examples, sometimes just one. How might this work?

This paper for this week explores that topic:

https://ai6034.mit.edu/wiki/images/LakeDec2015.pdf

As the paper indicates:

A central challenge is to explain these two aspects of human-level concept learning: How do people learn new concepts from just one or a few examples? And how do people learn such abstract, rich, and flexible representations? An even greater challenge arises when putting them together: How can learning succeed from such sparse data yet also produce such rich representations? For any theory learning (4, 14–16), fitting a more complicated model requires more data, not less, in order to achieve some measure of good generalization, usually the difference in performance between new and old examples. Nonetheless, people seem to navigate this trade-off with remarkable agility, learning rich concepts that generalize well from sparse data.

The paper explores this notion, proposes a mechanism, and demonstrates its performance on a number of problems. It's one of the more challenging papers we'll cover. Don't get lost in the math (there isn't much) or the mechanism (probabilistic programming); instead try to determine why this works. Why is the program successful? As usual there is no one right answer here, but see what you can come up with.

A old text describing the Spanish conquest of the Aztecs and Incas provides an obscure hint at one line of thought: "The Conquistadors on their Spanish Horses were seen as Centaurs, so at one were they with their horses during the conquest of Mexico and Peru." (If you're unfamiliar with the term centaur, check wikipedia.]

Note also that the very last page of the pdf has pointers to supplemental material. This is becoming more commonplace, as authors are asked to include more information to support their claims. Sometimes the supplemental material is raw data and code, sometimes it's pointers to additional papers/memos. Be sure to check out the material available so you'll know to do this in the future.

Week 10 -- November 22

With the skyrocketing interest in AI and machine learning has come a recognition that the models we create can be biased, sometimes by accident and at times even despite our best efforts. Given the application of these models to real world issues -- like being approved or turned down for a lown, or being grsnted parole or not -- these systems can have significant real world consequences.

This week we'll tiptoe into a quite deep and challenging subject: the fairness of algorithms.

We'll also proceed a little differently, with a two part exercise to be done before class.

Note that order here is very important: do #1 before #2. The papers in #2 will likely change your view of the answer to #1, but the whole idea is for you to think about this on your own, trusting your own intuitions on the subject, and grappling with the issue the way everyone has to.

Don't try to rewrite your answer to #1 after reading the papers; your answers to this part will be graded on whether you made a good faith effort to think about the issue on your own.

1) For the sake of concreteness, select either one of these systems:

a) a system that takes data about applicants and decides who to admit to a very competitive college (perhaps one not far from here that dates back to 1635)

b) a system that decides whether to approve a loan request

Describe what it would mean for either of these to be "fair." Try to be as explicit as you can, as a good definition would be computational, so it could be applied effortlessly to double check the outcome of the system. At the very least be explicit about what kinds of things make an algorithm fair or unfair. That is, what are your criteria for something being fair?

Make it at most one page long.

2) Read and comment on both of these papers in the usual 1-page summary. In the summary, please do not report what the papers say; assume I have read them (I have) and tell me what you think about what they say.

https://ai6034.mit.edu/wiki/images/A_Gentle_Introduction_to_the_Discussion_on_Algorithmic_Fairness.pdf

https://ai6034.mit.edu/wiki/images/Narayanan-183.full.pdf

In the Gentle Introduction paper pay particular attention to the different models of fairness. Does any one of them seem particularly appropriate or inappropriate to you?

In the Semantics paper, comment in particular on the reference in the abstract to a bias that is "veridical". What does that mean and why does it matter? (Yes, the second half of this question is vague, but you should be used to that from me by now. It means I want you to think about it carefully on your own.)

Just to be clear: this week you'll hand in two one-page writeups.

Week 11 -- December 6

AI and Ethics

For the ethics of AI we'll want to start with issues like, What does it mean to do ethical research in AI? What for that matter does it mean to do ethical research in any technology?

One good, easy to read source of guidance on this is available from the Markkula Center:

https://www.scu.edu/ethics-in-technology-practice/overview-of-ethics-in-tech-practice/

In particular, read at least these two sections:

1) Overview of Ethics in Tech Practice

2) Framework For Ethical Decision Making

Then read about two recent projects that produced controversy for Google:

The Maven project --

https://foreignpolicy.com/2018/06/29/google-protest-wont-stop-pentagons-a-i-revolution/

A search engine for China --

https://theintercept.com/2018/10/12/google-search-engine-china-censorship/

Answer these questions about those projects and the articles.

a) What actually is project Maven? Do you consider it unethical? Why or why not (feel free to use the framework in the Markkula Center materials).

b) What is the image at the beginning of the Foreign Policy story and how well is it matched to the actual content of the story?

c) Evaluate this claim from that article:

Officials stress that partnering with commercial industry on AI is a national security priority, particularly as potential U.S. adversaries such as Russia and China ramp up investments in that area. China, in particular, is dedicating $150 billion to AI through 2030, Floyd said. Meanwhile, the Defense Department is spending $7.4 billion on AI, big data, and the cloud in fiscal 2017, according to Govini.

d) It's easy to criticize Google's efforts to build a censored search engine for China, as there are numerous problems with it. But take the other side -- what possible benefits might come from it? (Serious answers only, please. "It'll make Google a lot of money" is not a serious answer, even if true.) In all ethical issues it's important (ethically!) to be able to see both sides of an issue. Ethics questions typically involve careful trade-offs and balancing acts. You have to be able to see both sides in order to judge the trade-offs.