Best To Dean Mynatt

IPaT page

Beth Mynatt is heading north to become the new Dean of Computer Science at Northeastern University. Georgia Tech will miss her; she has been a key part of Tech for over twenty years. Northeastern is getting a great leader, a valued colleague, and an excellent PhD graduate of Georgia Tech.

Today we hail her work on solving problems of aging and compare to what we do in theory. These musings wend toward solving problems of the kind raised in our previous post, on which we have an update from POPL General Chair Rajeev Alur that comes full circle to a Dean at Northeastern.

One aspect of aging is having a long memory of a field. When Ken and I were young, it was all about what humans can do with computers. Now it is much more about what computers can do with humans. With humans, not just for humans—the aspect of collaboration is key.

Aging…

Beth worked on various projects over her time at Tech. In 1999–2000, she was part of a team that launched the Aware Home Research Initiative (AHRI), “whose goal is to develop the requisite technologies to create a home environment that can both perceive and assist its occupants.” She has led the Institute of People and Technology (IPaT) since 2011, its inaugural year, where her team has helped to support numerous, impactful research programs for faculty across Georgia Tech.

To show how leadership and persistent work pay dividends, she is a co-PI on a new multi-institution grant led by Tech’s Sonia Chernova to build intelligent systems that support aging. The five-year, $20 million grant from the National Science Foundation, sponsored also by Amazon and Google, will go to create the NSF AI Institute for Collaborative Assistance and Responsive Interaction for Networked Groups (AI-CARING). The announcement says:

The institute aims to develop new longitudinal, collaborative AI systems that work with aging adults including those diagnosed with mild cognitive impairment, and their caregivers.

The AI-CARING front page says that they

“will develop a discipline focused on personalized, longitudinal, collaborative AI, enabling the development of AI systems that learn personalized models of user behavior, understand how people’s behavior changes over time, and integrate that knowledge to support people and AIs working together. These networked Human-AI teams will work with elderly adults and their caregivers in order to provide sustainable long-term care solutions.”

As someone who is not young, I can definitely see why this is important.

Missed the …

I sometimes feel that we in theory have missed the boat. The issue is that our problems are weighty—P=NP anyone?—but they do not directly impact practical computing. The areas that Beth is interested in, such as ubiquitous computing, have by definition had a large impact on real computing. The snippet from MIT’s Project Oxygen quoted by Wikipedia expresses the direction of impact:

In the future, computation will be human centered. It will be freely available everywhere, like batteries and power sockets, or oxygen in the air we breathe. … [C]onfigurable generic devices, either handheld or embedded in the environment, will bring computation to us, whenever we need it and wherever we might be. … We won’t have to type, click, or learn new computer jargon. Instead, we’ll communicate naturally, using speech and gestures that describe our intent … and leave it to the computer to carry out our will.

This may conjure a “Star Trek” vision, but some people already live with substantial parts of this reality. Is there a market for P=NP? On the “equals” side, we can think of a couple of movies featuring interested parties. For the “not equal” side, not so much?

If pressed to think of a theory result that “launched a thousand ships” of practical effort, Peter Shor’s theorem about factoring belonging to quantum polynomial time springs to mind. But the ship of universal quantum computing needed to get it under steam won’t come in for decades. It is quantum devices with rudimentary computational features that we see ruling in the meantime, while quantum communication protocols backed by quantum information theory are banked on now.

The Solutions Business

I, Ken writing from here, have often winced at the way the word “solutions” is used in business advertising. To a theorist, solutions are what make a conjecture become a theorem, what we grade in theory courses, what enjoyers of puzzles pursue for recreation.

I described a solution of the more practical kind in my “Pandemic Lag” post last July. It solves the problem of estimating the true current strength of rapidly developing young players (such as I was once) whose official ratings have been largely frozen for over a year and a half by the lack of in-person chess during the pandemic. Online play is not officially rated.

For example, if a preteen’s frozen rating is 1575, my formula will currently add 25 Elo points times 19 months of the pandemic to make 2050. Players with higher ratings and longer track records are adjusted less. In some recent instances, when such kids have defeated players with more-established ratings near 2200, people seeing the 1500s rating in the official tournament table have raised questions. My answer is based on the average performance of many children keen enough to compete in similar-level championship events. Thus my computer-intensive studies are safeguarding the welfare of minors.

Gender Gap and Pipeline in Chess

My rating solution raises another problem of the kind addressed in our last post. For a fresh instance, on Monday I submitted my final report on the European Team Championship. This ten-day tournament finished Sunday with the sensation of Alireza Firouzja becoming the youngest player ever rated over 2800, six months younger than world champion Magnus Carlsen was in 2009. (Carlsen will defend his title starting Friday in Dubai against the Russian Ian Nepomniachtchi.)

My intrinsic rating performance projections for the men’s/open section were all accurate to within 10 Elo points, within two-sigma error bars about . When restricted to the 32 juniors (out of 191 total players) whose ratings I adjusted, all projections were within 23 Elo points of the results and their average was within 2 Elo points—on adjustments averaging 80 Elo points per the 32 players.

In the women’s section, my gender-neutral formula adjusted up 39 of the 153 female players by an average of 169 Elo. The prescribed amount was larger because their ratings (average 2168) were lower than the 32 junior males’ ratings (average 2419) to begin with. The formula overshot their performance by 113 Elo, and accounted for the bulk of a 36 Elo average shortfall of my projections for the women’s section overall. This continues a pattern of observing not only a lower starting point, but also a lower first derivative when compared to males of the same rating, in tournaments with high junior participation where my adjustments are involved.

The Queen’s Gambit miniseries magnified awareness of the disparity—witness the article a year ago by the Australian economist and grandmaster David Smerdon, “What’s behind the gender imbalance in top-level chess?” My work offers a new way to pinpoint when and where the male and female pipelines diverge. I do not see how it could solve the disparity, however. Even converting my big spreadsheets of test results into publications is a tall ask—for one, they are considered sensitive data.

I have discoursed about the theory of my predictive model on this blog, and could say more about the thrill of empirical success in ways much of theory doesn’t reach. But this is still short of solving human problems in the manner of Dick’s intro, let alone solving the gender gap. On that, we’re grateful to have a communication from Rajeev Alur, General Chair of POPL, on how the statistics we noted came about and what the conference is doing.

More About POPL 2022

Rajeev began by observing that POPL PCs in previous years have been closer to the percentages we quoted for STOC and FOCS:

• 2021: 15.4% (8/52)

• 2020: 14.8% (8/54)

• 2019: 11.5% (6/52)

• 2018: 23.1% (12/52)

• 2017: 17.2% (5/29)

• 2016: 21.4% (6/28)

• 2015: 20.0% (6/30)

A more private figure that he gave us permission to divulge is that the proportion of PC invitations this year was just under 20% for women. Had they accepted at the same rate as the men, there would have been 12 women on the committee this year, as there was in the first year the PC size was expanded in 2018.

He went on to note an issue also raised in comments to our last post, namely that among the smaller population of women the same people are asked multiple times, leading to more declines. We can put the scaling problem another way: nearly doubling the committee size in 2018 also doubled the ratio of women being invited to their total number. Not all effects of scaling-up keep equal proportion. The root cause of course is not only the smaller population but its smaller first derivative: as Rajeev noted, the latest figures from the Computing Research Association in 2019 show that out of 417 PhDs in POPL-core areas, only 42 (10%) were female.

The main thing that they are doing is to take the kind of steps pointed up in the quotation from Valerie King in our post. As shown in the POPL 2022 Overview schedule on their front page, they have special lunch and breakfast events for women and LGTBQ attendees (the latter held since 2020) and mentoring workshops for graduate and undergraduate level students. POPL 2022 is the first to appoint a special Chair for organizing these events, Jennifer Paykin, who we note gave a special quantum-for-POPL presentation in 2020.

Next year’s POPL PC will be chaired by Amal Ahmed. Coincidentally, she is Associate Dean for Graduate Programs at Northeastern’s Khoury College where Beth Mynatt will be Dean. Alexandra Silva, who was on the POPL 2020 PC and on the POPL 2021 Organizing Committee as Accessibility Chair, is one of three Keynote Speakers this year. Hope for the pipeline picking up was raised by the 2021 SIGPLAN Robin Milner Award for Junior Researchers going to Emina Torlak of UW.

Of course, it will take a long time to translate the promotion of opportunity into closer parity. The low dip this year is what we noticed, but the long time effect is what our post highlighted at the end. Again we thank Rajeev for bringing both the time range and effort by SIGPLAN and POPL into greater context.

Open Problems

We’ve tried to “zoom out” where theory people are used to zooming in. How can that change perspectives on our field?

We are also thankful for how our field has branched out and wish everyone a happy Thanksgiving.

[fixed name in POPL section]