Friday, September 20, 2019

Changing Times

An old friend from college sent me an e-mail, and it got me thinking.  When I was an undergraduate at Harvard some significant number of years ago, I took the graduate level algorithms course offered by Michael Rabin and the graduate level complexity course by Les Valiant.  There were maybe a half dozen people in each of the classes.  (They were great classes, of course. But CS at Harvard back then was really, really small.) 

This semester, I'm teaching the graduate level course on randomized algorithms and probabilistic analysis.  Right now, the enrollment is 74 students;  well more than half are undergraduates.  Somehow, that says something to me -- about how the field has grown, and in at least some regards how Harvard has changed.  And about how much more prepared students are these days for these kinds of classes.  (Knowledge or probability is much more prevalent.)  Class sizes have been creeping up for so long that while it's noticeable year-to-year, it's much more stark and remarkable when I think back to my own time in college. 

Of course, it's also on my mind because it's a pain teaching a graduate class that large.  But it's a pain I can live with -- if I didn't like teaching, I wouldn't have become a professor.  And it's gratifying, if not a little bit shocking, that there's this kind of interest in the subject I really love, that I've been excited by for decades. 


Saturday, September 07, 2019

Off to ALGO/ESA 2019

I'm shortly hopping on a plane to head to ALGO/ESA.  I'll be giving a survey-ish talk on Learning Augmented Algorithms, covering my work so far in the area as well as some of the work by others.  I think it's a highly promising direction fitting in the framework of Beyond Worst Case Analysis, so I'm excited to give the talk, and hoping it's still a novel enough area to be new to most of the audience. 

For those of you who are there, feel free to say hi -- I'm looking forward to talking to people. 


Wednesday, September 04, 2019

Happy New Academic Year: Teaching Randomized Algorithms

It seems I haven't written on this blog for a while.

Today was the start of a new semester.  I'll be teaching Randomized Algorithms and Probabilistic Analysis, using the new edition of my book with Eli Upfal as a base, and throwing in other material.  (Everyone should buy the book!  Here's a link.) 

It's a graduate level class, but generally designed for first year graduate students, and there were a lot of undergrads "shopping" it today.  (We don't do pre-registration at Harvard, and students get the first week to choose classes, known as shopping.)  So many that people were standing out the doors of the room.  But because we have a bit of a shortage of classes this semester, I'm guessing there's a good fraction of students just checking it out.  We'll see Thursday, but for now I'll predict we'll fit in the classroom, and wait to see if I'm wrong.  (If I'm wrong, that's wonderful too.)   

It's been four years since I last taught the course, so this time I'm trying something new.  When I've previously taught the course, I tried to make the class inviting and friendly by telling the class we'd begin without assuming the class knew probability, and so the first couple of weeks would be reviewing basics (like, say, linearity of expectations and union bounds), albeit in a CS algorithms context.  This time, I let the class know I'm assuming they know (or will pick up) basic probability, and so they should read chapters 1-4 on their own, and we'll start with Chapter 5, Balls and Bins models.  Over the last decade, I've seen a huge shift in probability knowledge -- Stat 110, Harvard's probability course, has become one of Harvard's biggest classes.  Many students have already taking AI or ML or even data science courses where they've done some (further) probability.  It feels appropriate (and safe) to assume people entering in the class know probability, or can review what they need on their own, and start the class further along.

Now finally, a request.  It's actually hard for me to teach when using this book, because I don't want to just read the book to the students.  That's boring.  On the other hand, if I thought something was important, I most likely already put it in the book.  We have to mix up the standard lecturing format a bit.  So two things we'll be doing are

1)  doing some "puzzle problems" at the beginning of most classes, so people can try to solve problems.  (Kind of a flipped classroom approach, but not a full commitment.)
2)  reading papers, related to the class topics.

So if you have any good suggestions of probability puzzle problems, or readable papers (particularly application papers) that use relatively basic probabilistic analysis in neat ways, send them over.  I've got a semester to fill.

For curious people, here's one of today's starting problems, which I first learned about in graduate school.  (I'm pretty sure I owe thanks to Claire Kenyon for teaching it.  I'll link to the corresponding Wikipedia page on the problem maybe later.)

After lunch, Bob suggests the following game to see who pays.  Alice and Bob will each choose a different sequence of three flips.  (So they could choose "Heads-Tails-Heads'', or "Tails-Tails-Tails'' for example.)  After they choose, a fair coin will be tossed until one of their sequences appears as a consecutive subsequence of the coin tosses.  The player whose sequence appears first wins. (Note that if they choose the above sequences, and if the flips come up Heads-Tails-Tails-Tails, the player that chose Tails-Tails-Tails would win as soon as their subsequence appears;  it's not three flips, then start over again.)  Bob politely says that Alice can choose first, and after she chooses and tells him her sequence he'll choose a different sequence.  What should Alice choose?

Wednesday, January 09, 2019

ANALCO, SOSA, SODA post

I spent the last few days at SODA-ANALCO-ALENEX-SOSA in San Diego.  (Nice location choice, I'd say!)  Here's some news.

This will be the last ANALCO (Analytic Algorithms and Combinatorics).  Apparently submissions have been decreasing, so they've decided it will halt and the work on these topics will go into SODA and other conferences.  I'm not sure how to think of it -- I think we as a community have far too many conferences/workshops generally, but I think the SODA model of having ANALCO and ALENEX (and now SOSA, I imagine) folded in cleanly into the main conference is an excellent model.  I also like the ANALCO topics.  But I can understand the time may have come to do something else.  Thanks to everyone who worked to organize ANALCO and keep it going these many years.

It looks like SOSA (Symposium on Simplicity in Algorithms) will be taking its place in the SODA lineup.  I co-chaired the symposium with Jeremy Fineman this year, the second for the symposium.  I was surprised by the high quality of the submissions, and was then further surprised by the strong turnout at SODA.  The room was quite full for the Tuesday afternoon sessions, and there were easily 75+ people at several of the talks.  I do think there's a need for SOSA -- no other workshop/conference hits the theme of simplicity in our area, and it's a really nice fit with the rest of SODA.  I'm hoping it will last, and in particular that they'll continue to have a good number of high quality submissions, but that depends on all of you.  Ideally, there will be a positive feedback loop here -- now that there's a good home for this type of work (besides notes on the arxiv), people will be more inclined to write up and submit things to SOSA.  For Tuesday's talks, I'll call out Josh Alman's great presentation on "An Illuminating Algorithm for the Light Bulb Problem" as my favorite for the day.

With ANALCO exiting, though, I think there's more room for additional satellite events at SODA, so hopefully some people will get creative.

If I had thought about it I should have live-blogged the business meeting.  I'd say as highlights, first, Sandy Irani presented the report of the ad hoc committee to combat harassment and discrimination in the theory of computing community.   (See here for the report.)  There was an overwhelming vote to adopt their recommendations going forward.  It's good to see progress in addressing these community concerns.  Second, Shuchi Chawla will be the next PC chair, and she brought forward a plan to have SODA PC members be allowed to submit papers (with a higher bar) that was voted on favorably as well.

I suppose the last note is that Jon Kleinberg's invited talk was the conference highlight you expect a Jon Kleinberg talk to be, with interesting results and models related to fairness and implicit bias.

Thanks to SIAM and all the organizers for their hard work.