And Thanks for All the Fish, Altavista Version

All sorts of news about the plug finally being pulled on Altavista, which I still have an attachment to, being partially the product myself of DEC.  Here's a nice eulogy.  There's a good basic history at wikipedia's Altavista page. 

The book The Search: How Google and Its Rivals Rewrote the Rules of Business and Transformed Our Culture (mostly about Google, but covers other history as well) probably sums up Altavista's history as well as anything:

The mighty rise and fall with spectacular regularity int his business, and the pace of boom and bust only increased as the Internet took root in the mid-1990s.  Yet Altavista is remarkable for a number of reasons.  To borrow from the present, Altavista was the Google if its era.  In 1996, it was arguably the best and most-loved brand on the Web.  It presaged many of the current innovations and opportunities in search, from automatic language translation to audio and video search to clustering of results.  And as a business Altavista attempted -- and failed -- to go public three times in three short years under three different owners.  Possibly most instructive, Altavista was the product of a company that was an extraordinary success in its original business but ultimately failed because of hidebound management unwilling to drive by anything other than the rearview mirror. 

And Thanks For All the Fish

As my Area Administrator Tristen reminded me, "...today is officially your last day as my boss..."  Monday is July 1, which officially ends my term as Area Dean for Computer Science at Harvard.  The indefatigable David Parkes will be taking on the position.  (Thank you, David!  And my condolences!) 

While a ponderous exposition of all the wonderful things that have happened in Harvard CS is clearly called for, I'll try to keep it brief.  My main goal in taking the job was to turn our small group into a somewhat larger group, and I feel that has gone well.  We've hired five new excellent faculty over the last 3 years (Ryan Adams, Eddie Kohler, Jelani Nelson, Yaron Singer, and Stratos Idreos).  We've also done well in promotions, including multiple successful tenure cases, which was the other really important part of my job.  In other news, CS enrollments at Harvard are still booming, and while credit for that certainly belongs to others (a shout-out here to David Malan, who keeps bringing more and more students into our intro CS 50 course somehow), as Area Dean, I consider it my job to take credit for it.  (Similarly, while I'm at it, I'll take some credit for Les Valiant finally winning his long-deserved Turing award!)  Our faculty, who have always been friendly, cooperative, and worked together well continue to do so.  So I didn't break anything there (which is probably as good summary as any of my past three years).  My job was really to be a buffer with other administration so the rest of the faculty could go about their business being as great as they are.  And, as I've said, then taking a share of the credit for their greatness afterwards.  

I've already thanked all the faculty for putting up with me the last few years.  But special thanks goes to my Administator Tristen Dixey, who insists on calling me "boss" even though it's quite clearly more correct the other way around.  She makes CS at Harvard go.  And while all the faculty are always helpful, I very frequently leaned on the trio of Harry Lewis, Greg Morrisett, and Margo Seltzer for Area Dean advice, to make sure I didn't do anything too stupid.

Other thanks go to my graduate students -- both Zhenming Liu and Giorgos Zervas who previously graduated, and Justin Thaler this year -- for keeping me involved in (their) research.  (And all my other collaborators as well, but my students especially.)  Sorry you had to put up with me administrating while you were busy doing the work for graduating.   

It's hard to believe it's been three years.  I imagine someday I may find myself taking on another administrative position.  But for now, it's a nice feeling just to be done with this one.  

How Should We Choose Students?

Some of my previous posts have led me to think about the following -- something I'm hoping to write a longer piece about in the near future.

In the past few weeks, at Harvard (and elsewhere) there have been reports about the "decline of the humanities".  (Whether these reports have any significant bearing in reality is not necessarily important to this post.)  But machine learning keeps getting better and better.  While we may never be able to predict the exact outcome for an individual student, statistically speaking, as the universities gather more data, they will get better at predicting, for example, what a student will major in.  Potentially, with the right sort of tracking, universities may be able to predict reasonably well what jobs students may go into -- heck, they may get a statistically meaningful prediction of their future net worth.*  In particular, if we wanted to choose students according to what they were going to major in, in order to keep the humanities supporters happy, we could;  while we can already kind of do that now (based on, for example, what student say they want to major in), we'll just keep getting better at it.

This will lead to all sorts of questions.  Or, perhaps better said, will make the questions that already to some extent exist more pronounced.  First, getting to to the humanities concern, how should we choose our students?  Should we have quotas by future major?  We could assign departments a permanent percentage (well, an "expected percentage") of the incoming graduates and accept students accordingly?  From some faculty members' and administrators' point of view, perhaps this makes sense;  we can guarantee a department size, and a suitable faculty/student ratio per department.  To me, it seems potentially disastrous, turning the university into a static entity, which perhaps would not in any sense limit any individual student in terms of what they want to study, but would create a less flexible global atmosphere.  Again, in some sense, this question exists today;  at least some people have responded to the "humanities crisis" by saying that how students are accepted should be changed (to give preference to humanities-interested students), but the question becomes an even more significant challenge once you assume you actually have very strong prediction methods that can allow you to select students in this way more accurately than has been the historical norm.   

Of course, going beyond the picayune issue of whether we should choose students according to what they might major in, there's the larger scale question of how we should choose students.  Indeed, this question lies at the heart of many an affirmative action lawsuit, with the "reverse affirmative action" side claiming that people of what I will call "white" descent are not admitted in favor of less qualified "non-white" students.  (The issue is obviously more complicated than this paragraph can do justice to;  for example, the issue of Asian American discrimination arises.)  In such discussions, one generally hears the term "merit" -- if only schools just took the top people according to merit and ignored race completely -- but what exactly is merit?  Legislators or judges seem to want some sort of formula (usually based on grades and or test scores -- except that, by studying their own big data, some at Google claim that "G.P.A.'s are worthless" for their hiring).  Let's suppose our machine learning tools are good enough to estimate merit quite accurately if we define the merit objective function for them.**  How should we define it?  One particularly intriguing question, is the "merit" of the class simply the sum of merits of the collected individuals -- in which case we should ignore things like what major they want to choose -- or is the merit of the sum different from the sum of the merits?  I have some of my own not-completely-worked-out ideas, but again, this seems worth writing a longer essay about to work through the possibilities and implications.  

A further interesting question that arises is what sort of information can and should universities gather about applicants, in order to make these predictions.  College applications already ask for a lot -- grades, lists of activities, essays, letters of recommendation, test scores, sometimes interviews.  Suppose, though, that we could much more clearly predict your "merit" as a future student by parsing your Facebook account, or better yet, your e-mail from the last 3 years.  Should we be able to ask for that?  Perhaps we can guarantee that our algorithms will return a score only and your actual e-mail will not be examined at all by any human beings.  Or, by the time we get to the point where our machine learning algorithms are ready for that data, privacy won't matter to anyone anyway, especially if providing access to the data is needed to get into their choice of school. 

In some sense, none of these questions are inherently new.  But they appear to become different in kind once you think about the power machine learning will give to systems that make decisions about things like who goes to what university.  While the university setting is arguably small, the themes seem quite large, and perhaps the university is the place where some of the thinking behind the larger themes needs to be taking place.  And taking place now, before the technology is here and being used without a lot of thought into how it really should be used.

* Obviously, there are countless other potentially more significant uses of machine learning technology.  But I work at a university, so this is what has come to mind recently.   

** As far as I know, the merit function for Harvard is not "how much will you or your family donate to Harvard in the future".  But it could be.  Even if we avoid the potential self-interest of universities, to what extent is net worth a suitable metric of merit?  I was an undergraduate at Harvard and am now a professor there;  Bill Gates was an undergraduate (who notoriously dropped out) and donated a large amount of money for the building I now work in, and apparently has had a few other successes.  Extreme cases, to be sure, but how would the merit objective function judge these outcomes?  

Valiant's Book Out: Probably Approximately Correct

Les Valiant has a new book out: 
Probably Approximately Correct: Nature's Algorithms for Learning and Prospering in a Complex World

I was sent a free copy last week, but was delayed in reading it due to my avocational vocation.  (I was "talking with lawyers" a bunch.)  But I wanted to make sure to finish it over the weekend.  And now I'll recommend it all to you.

It would be, I think, somewhat inappropriate for me to attempt to review the book, but I'll aim to give some description of it which may encourage you to purchase it.  The book is aptly summarized by the following two sentences from it.

"The focus here will be the unified study of the mechanisms of evolution, learning, and intelligence using the methods of computer science."

"By the end of the book I hope to have persuaded the reader that when seeking to understand the fundamental character of life, learning algorithms are good place to start."  

Needless to say, the book is ambitious in scope, what one might expect from a Turing award winner, but in particular from Les.  If you have heard his Turing award lecture (available here), you can think of it as a preview of the book.  It is hard not to read the book as a challenge, to computer science in particular, but to the sciences more generally.  It is a call to arms, a vision, a plea, an agenda.

Because of this, I would recommend it highly to all computer scientists (in any area).

I would also recommend to it all scientists, so they could see this clearly laid out research vision from one of the leaders in computer science -- and, arguably, the one who is most interested in promoting the extension of the theory of computation to other sciences.  It might, I think, spur them to consider the relationship between computing and their own area of work, even if they are not directly working on evolution, learning, or intelligence.

It is slightly harder to recommend it to a general audience.  The book tackles fundamental questions of the connections between life and computation, making it a philosophical work certainly worthy of a large and general audience.  It raises some quite deep questions about the nature of human thought from what I think for most would be a novel vantage point.  But it does not shy away from the technical, and while, as promised, "The language of mathematics will be used, but only a little, and will be explained where used.", I imagine readers without a math/computer science background could get lost at times.  Still, other technical books (e.g., anything by Lisa Randall) find a large audience, so perhaps I underestimate the population at large. 

A final personal aside:  because I work with Les, when I read it, it came out in his voice.  I think the book very much sounds like Les -- it reads, to me, like him speaking -- but perhaps that's a trick of my own mind.  

Harvard Humanities

There's been a mild hubbub toward the end of the week here, due to a report and some articles (Boston Globe, WSJ) that the number of students majoring at the humanities at Harvard is in decline.  (See also this post at Shots in the Dark.)

Happily, this appears to be much ado about nothing.  Ben Schmidt at Princeton has already run the nationwide number, and shown that the decline is really more about a bubble in the 1960's of humanities majors.  Which just goes to show, when looking at historical data, what starting point you choose is important.  (Yes, that goes in the "duh", "lies, damn lies, and statistics" category.)

At Harvard, specifically, there are a variety of potential reasons for this trend, including but not limited to the general national trend.  In computer science, we've been actively trying to attract and retain students;  the humanities just may be facing more competition.  There is some claim that Harvard's financial aid policy is having an effect;  to the extent that students are coming from less well-off backgrounds, they may be seeking an education that they feel more directly will lead to job prospects.

There has been, however, perhaps a hint (or more than a hint) in some of all of what's going around that somehow people focusing on things outside of the humanities is "anti-intellectual", with students caring more about immediate job prospects than, well, the "intellectual" humanities. 

Naturally, I resent this.  I find computer science has a very solid intellectual basis.  The nature of computation, what it means to compute efficiently, how computing is found throughout nature (more on this in my next post) -- there's a lot interesting intellectually there.  If one seeks more "moral" sorts of lessons, I think many can naturally be found throughout CS, with the right interpretation.  The challenge of tradeoffs, for instance, is an underlying concept of my own algorithms class, and certainly appeared (if less quantitatively) in the moral reasoning class I took as an undergraduate. 

On the other hand, I understand where this is coming from.  There is a sense that the humanities is under siege (particularly at state institutions);  there are politicians of the mindset that "if it's not job training, why are we providing it?"  I believe that one should study more than computer science to learn to be a more complete human being;  I am thrilled to be at an institution where history, English, religious studies, as well as Romance languages, economics, and government are studied.  When one feels under attack, one's reactions might seem a bit more extreme.   

I'm not one to say where the final balance will be, or should be.  I do believe an understanding of computation should be a fundamental part of a liberal arts education;  it is clearly one of the most powerful ideas of the last century.  And it's our goal to make it both so that every Harvard student feels welcome and able to take a computer science course, and so that many understand our excitement and choose to major in it.  For a few decades, Harvard has been a bit behind in the role computer science has played at the university, and I think now that's changed.  So to the extent that the humanities feel the competition is from us, well, I'm actually all for it.    

Thanks to Harry Lewis for various discussions on this theme.  

NSF Reviewing Trial Run

Noam Nisan points to the NSF trying out some new rules for reviewing in its upcoming SSS program. 

There's a lot here to discuss.  First, I'm glad to see the NSF is willing to try out some new reviewing approaches.  They've been using the same approach for a long time now (1 or 2 day in person meetings, a reviewer panel drawn according to who is available and willing);  I really haven't seen any discussion from the NSF as to why it's a good review system, and it's typically got some major cons (as well as, admittedly, some pros).  But as far as I know -- and perhaps some people are more knowledgeable than I am on the topic -- it's not clear at all to me why it's become the stable equilibrium point as a reviewing method.

That being said, there's some clear pros and cons to this experiment.  Some features + initial off-the-cuff commentary.

1.  No panel review.  Proposals will be split into groups of 25-40, and PIs in the group will have to review other proposals (they say 7 here) in that group.  [If there are multiple PIs on a proposal, one has to be the sacrificial lamb and take on the role of reviewer for the team.]   

I kind of like the idea that people submitting proposals have to review.  One of the big problems in the conference/journal system is that there's minimal "incentive" to review.  Good citizens pay back into the system.  Bad citizens don't.  This method handles the problem in a natural way -- you submit, you review.  There are many potential problems with this method to be sure (as we'll see in the proposed implementation below).

2.  A composite ranking will be determined, and then the "quality" of the reviews of the PIs will be judged against this composite;  then the PIs ranking may be adjusted according to the quality of their reviews.

Ugh.  Hunh?  I get the motivation here.  You've now forced people into doing reviews, who may not want to.  So you need an incentive to get them to do the reviews, and do them well.  One incentive is that if you're late in your reviews, your own proposal will be disqualified.  That seems fine to me.  But this seems --- off.  I should note, they have a whole subsection in the document labelled

Theoretical Basis:

The theoretical basis for the proposed review process lies in an area of mathematics referred to as mechanism design or, alternatively, reverse game theory.  In mathematics, a game is defined as any interaction among two or more people.  The purpose of mechanism design is to enable one to “design” the “mechanism,” namely the game, to obtain the desired result, in this case to efficiently obtain high-quality proposal review while providing the advantages noted above.  In mechanism design, this is done by formulating a set of incentives that drive behavior in the desired direction.  The mechanism presented here was devised by Michael Merrifield and Donald Saari [¹].

I suppose I now have to go read the Merrifeld and Saari paper to see if they can convince me this a good idea.  But before reading that, there are multiple things I don't like about this.

a)   Why is "reviewer quality" now going to be part of how we make decisions about what gets funded?  I'm not sure to what extent, if any, I want "reviewer quality" determining who gets money to do research.  Here's what the document says:

To promote diligence and honesty in the ranking process, PIs are given a bonus for doing a good job.  The bonus consists of moving their proposals up in the ranking in accordance with the accuracy with which their ranking agrees with the global ranking.  This movement will be sufficient to provide a strong incentive to reviewers to do a good job, but not large enough to severely distort the ranking merely as a result of the review process.  Recognizing that, if all reviewers do an excellent job of ranking the proposals they review, all PIs’ proposals will be moved up equally, which means that the ranking will not be changed, the maximum incentive bonus will be a movement of two positions, that is, a proposal could be moved up in the ranking to a position above the next two higher proposals.

With funding ratios at about 15% (I don't know what the latest is, but that seems in the ballpark), two places could be a big deal in the rankings.  

b)   Why is there the assumption that the group ranking is the "right" score -- particularly with such small samples?  I should note I've been on NSF panels where I felt I knew much better than the other people in the room what were the best proposals.  (Others can judge their confidence in whether I was likely to have been right or not.)  One of the pluses of face-to-face meetings is that a lone dissenter has a chance to convince other reviewers that they were, well, initially wrong (and this happens non-trivially often).  I'm not sure why review quality is judged by "matching the global ranking".

c)   Indeed, this seems to me to create all sorts of game theoretic problems;  my goal in reviewing does not seem to be to present my actual opinion of a paper, but to present my belief about how other reviewers will opine about the paper.  My experience suggests that this does not lead to the best reviews.  The NSF document says:

Each PI will then review the assigned subset of m proposals, providing a detailed written review and score (Poor-to-Excellent) for each, and rank order the proposals in his/her subset, placing the proposals in the order which he/she thinks the group as a whole will rank them, not in the order of his/her personal preference.

But then it says:

Each individual PI’s rankings will be compared to the global ranking, and the PI’s ranking will be adjusted in accordance with the degree to which his/her ranking matches the global ranking.  This adjustment provides an incentive to each PI to make an honest and thorough assessment of the proposals to which they are assigned as failure to do so results in the PI placing himself/herself at a disadvantage compared to others in the group.

So I'm saying I'm not clear myself how their incentive system -- based on the global ranking --- gives an incentive to make an honest and thorough assessment.  Even the document itself seems to contradict itself here.

d)  This methodology seems ripe for abuse via collusion -- which is of course against the rules:

The PIs are not permitted to communicate with each other regarding this process or a proposal’s content, and they are not informed of who is reviewing their proposals.

But offhand I see plenty of opportunities for gaming the system....

e)  This scheme is complicated.  You have to read the document to get all the details.  If it takes what seems to be a couple of pages to explain the rules of the assignment and scoring system, maybe the system is too complicated for its own good.

That came out pretty negative.  Again, I like the idea of experimenting with the review process.  I like the idea that submitters review.  I understand the concept that we somehow want to incentivize good reviews, and that's very difficult to incentivize.

This actual implementation... well, I'd love to hear other people argue why it's a good one.  And I'd certainly like to hear what people think of it after it's all done.  But it looks like the wrong way to go to me.  Maybe in the morning, with some time to think about it, and with some comments from people, it will look better to me.  Or maybe, after others' comments, it will seem even worse.