Things I’ve Learned by/about Playing Roller Derby

So, uh, I play roller derby. Recently, I wrote an essay outlining some of the things that I’ve learned by/about playing roller derby over the past couple of years. I originally posted this essay on my league’s forum, but the more I think about it, the more I believe that many of these points apply (albeit in some modified form) to academia. So I’m also posting a link to it here too. Enjoy!

The President on Probability

Barack Obama on reasoning under uncertainty:

“Nothing comes to my desk that is perfectly solvable,” Obama said at one point. “Otherwise, someone else would have solved it. So you wind up dealing with probabilities. Any given decision you make you’ll wind up with a 30 to 40 percent chance that it isn’t going to work. You have to own that and feel comfortable with the way you made the decision. You can’t be paralyzed by the fact that it might not work out.” On top of all of this, after you have made your decision, you need to feign total certainty about it. People being led do not want to think probabilistically.

(From this article.)

ML Hipster

*sigh* I have a total nerd crush on ML Hipster.

(Via Chuck.)

A Letter of Appreciation

Ten years ago, I was an M.Sc. student at the University of Edinburgh, about to start my thesis research on “FSA Induction for Real World Datasets.” A little over ten years ago to the day, Miles Osborne, my thesis advisor, emailed me to say that there was a talk taking place the next day that I absolutely had to attend: “From Grep to Graphical Models” by Fernando Pereira from the University of Pennsylvania. As someone whose academic interests included UNIX tools, regular expressions, finite state automata, HMMs, and graphical models, my mind was blown by the title alone.

I attended the talk and I loved every second of it. Fernando introduced conditional random fields, a new probabilistic model for sequential data that he and two other researchers, Andrew McCallum (or “??? from WhizBang” according to my handwritten notes!) and John Lafferty, had just invented. Sure, I didn’t understand everything he said, but to me, CRFs sounded like the most interesting thing on earth: not only did they draw on ideas from HMMs and multiclass logistic regression (which I’d recently learned about in Miles’ class on “Data Intensive Linguistics”), CRFs were undirected graphical models, unlike the (directed) graphical models that I’d previously encountered. At the end of Fernando’s talk, I told Miles that I wanted to do my thesis research on CRFs. Miles laughed and said he’d been planning to convince me to work on CRFs anyway, so this switch was perfect.

Miles arranged for me to meet with Fernando. I still have my handwritten notes from that meeting. I learned about all kinds of things that, at the time, I only half—no, maybe one quarter—understood. Each one of them sounded unbelievably exciting to me. When I told Fernando that would be starting a Ph.D. at the University of Cambridge after my M.Sc., he asked me why I hadn’t applied to Penn. My response (which I’m still embarrassed by!) was that I’d never heard of Penn. He told me that it wasn’t too late to change my mind, but I was firmly convinced that Cambridge was the center of the universe and, besides, I’d never heard of this no-name university, Penn.

I spent the rest of my M.Sc. working on CRFs. I spent months poring over the original CRF paper. I rederived every single equation myself. I tracked down books and papers from the ’70s and ’80s about Markov random fields, the Hammersley Clifford theorem, and semirings. I downloaded statistics articles from JSTOR. I read machine learning papers on everything from factorial HMMs to numerical optimization methods. It was an awesome experience (and ultimately resulted in this).

At the end of my M.Sc., I moved to Cambridge and started my Ph.D. Three months later, in December, I attended my first NIPS conference. I had one goal: to talk to all three authors of the CRF paper and to convince one of them to let me work with them over the following summer. I achieved my goal. I ran into Fernando on a staircase in the Westin, I introduced myself to Andrew McCallum after a talk, and I cornered John Lafferty (poor man!) in an elevator. All three said yes.

On Christmas Day (or maybe Christmas Eve?) I received an email from Fernando saying he’d just read my M.Sc. thesis and had some questions. (To cut a long story short, due to my lack of a statistics/probability background, I’d made a “thinko” regarding linearity of expectation. Oops.) I was amazed and honored that he’d read it—and that he’d done so over his Christmas break.

Over the next few months, I received emails from both Fernando and Andrew; however, as a disorganized Ph.D. student, unsure of what I wanted to do, I don’t think I replied. Eventually, in May, I received a contract from Fernando indicating that I would be spending three months at Penn, starting in June. (We may have exchanged emails prior to that, but I’m not sure.) I signed it.

On June 19, I flew from London to Philadelphia. I took the SEPTA train from the airport to the Penn campus, where I single-handedly lugged my giant suitcase up a flight of stairs in the hottest weather I’d ever experienced. I somehow managed to make my way to the CS department (this was back in the days before GPS and smartphones etc.), where I found Fernando in his office.

I’m not really sure what to say at this point because, really, that was when my life as an adult began. Spending that summer at Penn was the biggest adventure of my life. And I loved every second of it. Okay, sure, I got mono and spent a month incredibly sick and barely able to move, but apart from that, it was amazing. At the end of the summer, I didn’t want to return to Cambridge.

To cut a long story short, Fernando offered to let me stay at Penn for a year. Somehow, one year turned into four, and I ended up doing a strange, bicontinental Ph.D. Fernando paid my salary and paid for my trips back to Cambridge to satisfy my Ph.D. requirements. I’m still not entirely sure what he got out of it—I never co-authored any publications with him and his name didn’t end up on the cover of my Ph.D. thesis. In fact, for most of the time I was at Penn, I wasn’t even working on topics that aligned with his research agenda. And yet, he was unfailingly supportive. He funded most of the first WiML workshop. He met with me regularly and gave me advice on my research, as well as academia in general. And, most importantly, he believed in me even when my confidence was at an all-time low. I am indescribably grateful to him for his generosity. I’m not sure I will ever be able to convey the extent to which his generosity changed my life and made me who I am.

Ten years later, I’m an assistant professor. I’ve worked with two of the three authors of the CRF paper. (When I left Penn, I did so to do a postdoc with Andrew McCallum.) The WiML workshop is now in its seventh year. I do research on topics that I love. But most importantly, I hope that one day I’m able to be as good a mentor to, well, ANYONE as Fernando was—and still is—to me.

TL;DR: Thanks, Fernando. You rock.

On Being Oneself

Tim sent me a link to this blog post, which consists of a quote from this blog post. I think the quote is well worth taking seriously, especially if one is a Ph.D. student, postdoc, or young faculty:

[A] young person told me that I could hold to my principles about the importance of my family, honesty and equality—and any of a hundred other things because I had “made it.” This troubled me. It troubles me when I hear the same thing from new Ph.Ds who are trying to get tenure. I don’t see how you can pretend to be someone else for 5 or 10 years until you have “made it” and then be your true self.

Videos on Computational Social Science

For the past two years, the UMass Amherst Computational Social Science Initiative has been running a weekly seminar series. We’ve had some amazing speakers: thirty-one of them to be precise (though I may have miscounted). But here’s the really exciting bit: we videoed twenty-three of the talks and the videos are available online. As far as I’m concerned, this is an unbelievable resource for anyone interested in computational social science and, as someone who was present at almost every one of these talks, I can tell you they definitely worth watching.

Women in CS at Harvey Mudd (Again)

Another article about Harvey Mudd’s Maria Klawe. (Previously disucssed here.)

How to Be a Successful Ph.D. Student

Mark Dredze and I have written a guide on “How to Be a Successful Ph.D. Student.” It’s a work in progress, so we’re looking for feedback from everyone: those who were/are Ph.D. students, and those who advise students. You can either post feedback here or email me directly.

Write First

Recently, I’ve been trying to follow two complementary pieces of advice on paper-writing. The first, which I’ve known about for some time now and disucssed with my research group back in September, is Jason Eisner’s “Write the Paper First.” Jason advocates going against the trend of last-minute paper-writing commonly found in computer science. He provides many well-argued reasons for this viewpoint, but one of the most compelling (to me, at least) is the following:

But you can’t write effectively [on little sleep]. Writing involves many big and small decisions, which will seem insurmountable when you’re exhausted and panicked.

This observation ties in especially well with recent research on decision fatigue.

The second piece of advice is George Whitesides’ “Writing a Paper.” Whitesides (who has the highest h-index of any living chemist) argues for the continual use of outlines when writing papers:

A paper is not just an archival device for storing a completed research program; it is also a structure for planning your research in progress. […] A good outline for the paper is also a good plan for the research program. You should write and rewrite these plans/outlines throughout the course of the research. […] The continuous effort to understand, analyze, summarize, and reformulate hypotheses on paper will be immensely more efficient for you than a process in which you collect data and only start to organize them when their collection is “complete.”

Of particular relevance to advisors and students is the following paragraph:

An outline […] contains little text. If you and I can agree on the details of the outline (that is, on the data and organization), the supporting text can be assembled fairly easily. If we do not agree on the outline, any text is useless. […] It can be relatively efficient in time to go through several (even many) cycles of an outline before beginning to write text; writing many versions of the full text of a paper is slow.

WiML Call for Executive Board Members

Posting in my capacity as the current chair of the WiML Executive Board:

We are seeking new members to join the Women in Machine Learning (WiML) Executive Board. The goals of the Executive Board include ensuring the continued success of the annual WiML Workshop, facilitating activities with a lifespan longer than one year, securing funding, and handling publicity and communications. In previous years, the Executive Board has secured grants to fund the Workshop, analyzed impact statistics, and established a mentoring program, as well as advising the Workshop organizers and maintaining infrastructure.

Prospective Executive Board members must

  • commit to a two year position,
  • have participated in at least one WiML Workshop,
  • demonstrate active participation in the machine learning community,
  • have experience/interest in broadening women’s participation in CS.

Senior Ph.D. students, postdoctoral researchers, faculty, and research scientists (in both industry and academia) are encouraged to apply.

Applicants must submit a short (500 words or fewer) statement of interest and suitability, to include the following information:

  • Affiliation,
  • relevant previous experience,
  • reasons for interest and suitability.

Statements should be sent to wimlworkshop@gmail.com by March 30, 2012.