New Public Quantum Network in Urbana

My local library has just booted up a public access point to a quantum network. 

Huh?  What?

This appears to be the first public access quantum network on the planet!  Neat.

I’m very proud of my little town and our library.   This is yet another ground breaking technical exploration at this library. In the past, the library has pioneered public access to the Internet (which we invented here) and a decade ago I helped boot up a teen lab featuring 3D printing and other personal fabrication tools. 

Essentially, there is a fiber link between The Urbana Free Library and a physics lab at the UIUC, a couple of kilometers down the street.

This demo network is intended for anyone in the community to experiment with.

Which begs the question, what in the world should we do with it?  What can we do with it?

I’m still trying to figure this out.  (Part of the challenge is that this is just networking, it’s not quantum computing.)

The initial demonstration is pretty simple and there isn’t much you can do with it.

I gather that Real Soon Now there will be quantum encryption and messaging.  I’ll have to see what the public interface will be.  I mean, we already have various kinds of messaging, including encrypted messaging.  (Not that I need to send a message from the library to the physics building very often.) 

So, we’ll have to see.

---

Some questions to think about for the future:

How can I connect my regular devices to this quantum link?  Do we need to create some kind of bridge?  I’m sure we can do that, if we can get permission.

If I can connect to this link, can I use the encryption to create digital objects, e.g., digital signatures?  If so, and I can use these objects in regular software, there are tons of demonstrations to do.

Thinking about it, what I want to do is put this encrypted, secure, link in between client and server systems, and between multiple servers, or multiple apps.

Anyway, congrats to TUFL, UC2B, and Illinois Quantum.

Why did we do this?

Because we can!

What Would You Do With Four Extra Hands?

Just down the road from me, Professor Kim and his happy elves in the Kimlab have been fiddling around with robot arms.  This fall they demonstrated the PAPRAS Backpack, which is a “Robotic Backpack System with Pluggable Supernumerary Limbs” [1].

Clearly, the main purpose was to show that, “yes, yes we can build it!”  And one of their videos is explicit:  this is inspired by comic books and movies. (A 2021 video was titled “PAPRAS: Backpack (Tribute to Dr. Octopus in Spider-Man)”

Fair enough.

But, I have to say that this demonstration raises more questions than it answers.  How would you ever use something like this?  What would you do with it?  Where do you get shirts with six arm holes? 

Even in the Hollywood version, the extra arms seem more decorative than useful.  (Not to mention that the extra arms seem to be symptoms of Dr. Octopus’ serious mental issues.)

Clearly, controlling four extra limbs is going to be a challenging mental load, and will take quite a bit of practice.  One also wonders about safety protocols, with your reach extending in radically unfamiliar ways. So, there is lots of interesting research to be done here.

If you assigned me to teach someone to safely and sanely control this backpack, and to do useful things with it, then I think I would want to cooperate with experts in embodied movement—dancers.  I’m pretty sure that there are some folks very near the Kimlab who would be very interested to collaborate.

Anyway, this is a great student project.  Give them an ‘A’!

---

1. Chaerim Moon, Sean Taylor, Kevin G Gim, Sankalp Yamsani, Kazuki Shin, Kyungseo Park, and Joohyung Kim, Robotic Backpack System with Pluggable Supernumerary Limbs, in EEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2023). 2023: Detroit.

Robot Wednesday

New Public Art Nearby

Like many public institutions, the University near me has a longstanding “1% for art” policy, which has populated the campus with hundreds of sculptures and other art works [2].  Many of them are forgotten, even in plain sight.  But a few have been remarkable successes.

I have written before about the important role the Alma Mater has acquired [1].  What ever the intentions of the artist and administrators 100 years ago, Alma has become the focus of a happy communal ritual that unites the whole University community at graduation time.  Remarkably, this ritual was spontaneously created by the students, without official directions or “permission”. 

This spring, a new art work was installed on campus, and it is already a great success.  Untitled, the installation is a full size model of a Mammoth, which might have lived in the area 6,000 years ago [3]. Prominently located near the historic nineteenth-century Natural History Building, it has been noticed by the thousands of daily passers by!

The powers-that-be have their own official story of what this work might mean.  It is attention grabbing (and pretty), and also “giant educational opportunity.”  This is a highly realistic representation, evocative of the cherished Natural History Museum, and the long tradition of natural science on campus. 

But, of course, within weeks, the student body seized on this art work, and is creating their own meaning. (Again, without official instruction or “permission”.)

First of all, she (?) has been given a affectionate name:  Tuskee-Wow-Wow.  If you are in anyway familiar with Illinois, you recognize the reference to the enigmatic school song.  I, for one, was immediately charmed by this moniker!

Furthermore, and fueling my own suspicion that it’s a “she”, Tuskee-Wow-Wow was incorporated into the annual Mother’s Day weekend, with an informal dress up [4]. 

(From [3])

I expect she wlll regularly join Alma, who dresses up for many occasions.

I say, welcome to campus, young Mammoth!  We love you already.

---

1. Robert E. McGrath, A Digital Rescue for a Graduation Ritual. Urbana, Illinois, 2019. http://hdl.handle.net/2142/105503
2. Muriel Scheinman, A Guide to Art at the University of Illinois: Urbana-Champaign, Robert Allerton Park, and Chicago, Urbana, University of Illinois Press, 1995.
3. Travis Tate, A mammoth tribute to Illinois history, in College of Liberal Arts & Sciences – News, March 24, 2023. https://las.illinois.edu/news/2023-03-24/mammoth-tribute-illinois-history
4. Bradley Zimmerman, Bright Spot: U of I woolly mammoth decorated for Mom’s Weekend, in WCIA TV News, April 14, 2023. https://www.wcia.com/news/bright-spot/bright-spot-u-of-i-woolly-mammoth-decorated-for-moms-weekend/

Bardeen’s Demo Box

The transistor is 75 years old, which means that the rigged demo must be almost that old!

After John Bardeen helped discover “the transistor effect”—which merely revolutionized everything and paved the way for digital computers and networks—he was recruited to the University of Illinois.  He stayed the rest of his career, messing about in the lab and working out superconductivity a second Nobel.  (I was told by a professor that the University’s policy was to keep him here no matter what.  Chain him to the EE building, if necessary. : – ) )

Many thousands of us went to school here, inspired to build on the achievements of Bardeen’s generation.  We produced many, many rigged demos over the years to show what we had learned.  It’s the way we do it.  (I made a joke version of one of my papers, to make it read, “in partial fullfillment of a Master’s Degree in Demonstration Science”.)

This fall Allison Marsh reports on the demo that was created to show off this new thing, the transistor, which is in a museum now [1].  It was called “Transistor Oscillator-Amplifier Box” (the title needs work!), and it beeped.  And, as several generations of students know, you can make it beep a tune.  The standard demo is “How dry I am.” Bardeen’s handwritten note is attached.

It’s not how great the beeps are, it’s that it beeps at all!

I encountered this box as a kid, and it was always an inspiration.  I actually wrote a memo to the Spurlock Museum, suggesting that it should not just sit in a glass case, there should be a multimedia station that plays the sounds.  (At the time, we were consulting with the museum on possible high tech augmentations.)

I was very happy to see that the box has had some serious conservation work done, and is now better documented including some digital recordings of the sounds.

It’s good to have the video, but I still want an interactive display in the musuem.  The original cannot be demonstrated, but a recording and sound system could faithfully simulate the experience. 

We’ll see what happens. Museum time moves way, way slower than digital technology time.  They aren’t going to rush into anything, no matter how great my ideas may be!

Anyway, at this 75^th anniversary of a truly revolutionary technology, we should acknowledge the giants who walked among us.  “We are not worthy!  We are not worthy!”

And we who follow will keep playing around in the lab, coming up with amazing ideas and demonstrating that they really work. 

---

1. Allison Marsh, John Bardeen’s Terrific Transistorized Music Box, in IEEE Spectrum – History of Computing, November 30, 2022. https://spectrum.ieee.org/transistor-music-box

Generalized Automated Chemistry At UIUC

Automating chem labs is only the first step.  Once we our workflows include robots, we can unleash machine learning to optimize the process.

This fall researchers down the street at the University of Illinois Urbana Champaign report, and I quote:

“a simple closed-loop workflow that leverages data-guided matrix down-selection, uncertainty-minimizing machine learning, and robotic experimentation to discover general reaction conditions.”

([1], p.1)

Or, as the press release puts it, “powerful AI and a molecule-making machine to find the best conditions for automated complex chemistry”. [2]

If I understand correctly, the point is to discover general processes for synthesizing a desired molecule.  There are many starting places, and many paths, so a brute force search is difficult.  In practice, chemists usually start from a convenient input (e.g., from a small set of commonly used conditions), which works but is not optimal for most purposes. 

The goal of this research is to figure out a better set of starting places.

The researchers report that they tried to use machine learning on the literature, seeking optimal starting points.  This effort failed for the interesting reason that negative results are generally not reported.  The ML could not learn without these results.

(!)  

We’ve been telling you to publish negative results, and now we have a really good reason: they are useful!

The logical thing to do, then, is to connect the ML with a robot lab, and let it generate its own results, successful and not.  At least if you work at my old Alma Mater, that’s the logical thing to do. : – )

Basically, this games is a search for short, fast, and cheap paths through a ginormous space of possible chemical building blocks and ways to combine them, looking to get to the desired molecule.

First, the system used published data to guide the search by “down-selecting” the starting points, i.e., picking out a small, strategic set of good places to start.  The selected set were used in “seeding” experiments to further prune this set into “only” 512 initial conditions.

Second, machine learning was used to guide the search.  This apparently involved some non standard data wrangling and probabilistic reasoning; not-quite-Bayesian plus “active learning”, minimizing uncertainty.  (Honestly, I don’t know this math at all. I’m taking their word for it.)

Third, let’s close the loop.  The search above is marching through samples of experimental conditions.  So, at each iteration, the robot executes the experiments, and returns the results to the ML model.  (The system prioritized experiments that were “unexplored”.) After only a relative few rounds, the uncertainly was not changing,

The result is a list of initial conditions and predicted yields. The ML actually explored a range of yields.  The top yields known from the literature were near the top, but there were also reactions with even higher yields than previous reports.

Cool!

The researchers note that the dataset includes a wide range of results, showing that the search “learned by probing both low- and high-yielding conditions”, unlike the published literature which is “heavily skewed toward positive outcomes”. ([1], p.6)

Looking at the progress of the model, it seems to look for good reactions first, then looks wider for other even better candidates, and later focusses on reducing uncertainty by exploring “negative results”.

The resulting conditions are “higher-yielding general conditions”, i.e., good places to start for many different reactions.  The model’s selections were tested as inputs to generate a variety of molecules.  The results were substantially higher yields than the standard benchmarks, with the top achieving double the benchmark value.  In addition, more of the samples achieved a practical minimum yield, i.e., would be useable.

Nice work all. Very neat stuff!

---

1. Nicholas H. Angello, Vandana Rathore, Wiktor Beker, Agnieszka Wołos, Edward R. Jira, Rafał Roszak, Tony C. Wu, Charles M. Schroeder, Alán Aspuru-Guzik, Bartosz A. Grzybowski, and Martin D. Burke, Closed-loop optimization of general reaction conditions for heteroaryl Suzuki-Miyaura coupling. Science, 378 (6618):399-405, 2022/10/28 2022. https://doi.org/10.1126/science.adc8743
2. Liz Ahlberg Touchstone, Artificial intelligence and molecule machine join forces to generalize automated chemistry, in Illinois Research News, October 28, 2022. https://news.illinois.edu/view/6367/1723467564

Common Workflow Language is Good Stuff

I’ve been following the challenge of reproducibility for a long time.  Inevitably, this mountain-we-have-to-climb involves both “provenance” and “workflows”.  The former is metadata describing how a computation happened in enough detail to understand it, and the latter is technology for describing how to do a computation, including necessary non-digital operations.

The first goal is for these descriptions to be portable across time and contexts, so they can be shared, evaluated, and so computational results can be trusted (or at least understood).

But obviously, these technologies are closely related.  Ideally, a description of the provenance of a digital artifact should be “executable”, i.e., you should be able to generate a workflow that reproduces all the steps (and, ideally, the results).  Conversely, executing a workflow should generate metadata about what happens, which together describes the provenance of the results. It should be a circle, no?

In the last couple of decades, quite a bit of technology has been developed in these areas.

If anything, too much technology has been developed, in that there are lots of similar, but incompatible ways to solve these problems. 

Which isn’t ideal. 

If your cool workflow only works on your system, but not on mine, that’s not really what we are looking for.

What’s the solution?

Another level of indirection, of course!

---

In this case, a new standard for interchange of workflows, the Common Workflow Language [1].

(I can tell this is both serious and well thought out because Sensei Carole Goble is involved.)

The basic idea is to create a neutral language for describing a computation, which specific tools can both read an write.  I.e., I can save my workflow in CFL, and you can read the CFL into your tool.

The standard builds on decades of experience with workflows and workflow description languages, and is specifically designed for portability across contexts.

The CFL includes languages for describing computational tools (e.g., what are the controls and options available), and also has a model of execution, i.e., running a computations.  The latter necessarily includes concepts of data flow, and an abstract description of data objects.

I can tell that this crew knows what they are doing, because they include a section, “When is CWL not useful?”  ([1], p. 61)

In general, “workflows” work for “batch” computing, and don’t work for things that depart from the “start at the beginning, proceed through the middle, and come out at the ending” model. 

Basically, computations that depend on complex conditions and/or external conditions (e.g., a specific day and time) are difficult to describe and difficult or impossible to reproduce. So, as the authors summarize it, CWI doesn’t work for

• “Safe interaction with stateful (web) services.
• Real-time communication between workflow steps.
• Interactions with command-line tools beside 1) constructing the command line and making available file inputs (both user-provided and synthesized from other inputs just prior to execution) and 2) consuming the output of the tool once its execution is finished, in the form of files created/changed, the POSIX standard output and error streams, and the POSIX exit code of the tool.
• Advanced control-flow techniques beyond conditional steps.
• Runtime workflow graph manipulations: dynamically adding or removing new steps during workflow execution, beyond any predefined conditional step execution tests that are in the original workflow description.
• Workflows that contain cycles: “Repeat this step or sub-workflow a specific number of times” or “Repeat this step or sub-workflow until a condition is met.”
• Workflows that need specific steps run on a specific day or at a specific time.”

([1], p.61)

The CWI is not the first rodeo for this group, so the standard does all the important things including an open source set of tools and plug ins so that you can use CWI in your favorite working environment.  I cannot overstate the importance of such tooling.

Overall, this is really useful stuff, well designed and well executed.  It solves a bunch of important problems in a really good way.   I hope it conquers the computational world.

---

1. Michael R. Crusoe, Sanne Abeln, Alexandru Iosup, Peter Amstutz, John Chilton, Nebojša Tijanić, Hervé Ménager, Stian Soiland-Reyes, Bogdan Gavrilović, Carole Goble, and The CWL Community, Methods included: standardizing computational reuse and portability with the Common Workflow Language. Communications of the ACM, 65 (6):54–63,  2022. https://doi.org/10.1145/3486897

Report on Local Pandemic Management At U. Illinois

Where I live we had a significant challenge dealing with the COVID pandemic:  a major university campus.  This institution high density residences and lots of personal interactions, plus a very mobile population.  Unchecked, the virus would quickly infect all the student, teachers, and staff, and then spread to the surrounding populations.  Before vaccines and treatments were available, this was a potentially explosive situation.

Our University stepped up with a technically advanced and well-designed program of rapid universal testing, contact tracing and isolation, and, of course, remote instruction and meetings.  And later, with nearly universal vaccination.  (Read the research paper for details [1].)

So, how well did this work?

This summer researchers report estimates of the effectiveness of these measures [1].   The bottom line is that it looks like these measures worked pretty well.  Millions of tests were performed—tens of thousands per day—yielding very low positivity.  I.e., nearly all cases were detected, and detected early.

Statistical analyses suggest that the area overall experienced substantially fewer cases than similar places with less effective management.  It’s hard to say, but clearly, things would have been a lot worse if the university population had been a major transmitter of the virus throughout the area.

Living in the region, I have to say that I was very pleased with the aggressive and effective measures at the university.  I was extremely worried that the university would be a major site of infection and would spread quickly not only locally but also back to the homes of the students as they come and go.  For me, it was a great sense of relief when it became clear that the university was doing an outstanding job of protecting the rest of us. We still had to close things down, mask up, and suffer. But at least we weren’t defeated by uncontrolled transmission from campus.

The university wants to crow about this success, calling it “a model for effective pandemic management” [2]. I don’t always agree with the my Alma Mater’s self-congratulations, but in this case,  from my perspective, this accolade is well justified.

Very nice work, all. Thank you.

---

1. Diana Rose E. Ranoa, Robin L. Holland, Fadi G. Alnaji, Kelsie J. Green, Leyi Wang, Richard L. Fredrickson, Tong Wang, George N. Wong, Johnny Uelmen, Sergei Maslov, Zachary J. Weiner, Alexei V. Tkachenko, Hantao Zhang, Zhiru Liu, Ahmed Ibrahim, Sanjay J. Patel, John M. Paul, Nickolas P. Vance, Joseph G. Gulick, Sandeep Puthanveetil Satheesan, Isaac J. Galvan, Andrew Miller, Joseph Grohens, Todd J. Nelson, Mary P. Stevens, P. Mark Hennessy, Robert C. Parker, Edward Santos, Charles Brackett, Julie D. Steinman, Melvin R. Fenner, Kirstin Dohrer, Michael DeLorenzo, Laura Wilhelm-Barr, Brian R. Brauer, Catherine Best-Popescu, Gary Durack, Nathan Wetter, David M. Kranz, Jessica Breitbarth, Charlie Simpson, Julie A. Pryde, Robin N. Kaler, Chris Harris, Allison C. Vance, Jodi L. Silotto, Mark Johnson, Enrique Andres Valera, Patricia K. Anton, Lowa Mwilambwe, Stephen P. Bryan, Deborah S. Stone, Danita B. Young, Wanda E. Ward, John Lantz, John A. Vozenilek, Rashid Bashir, Jeffrey S. Moore, Mayank Garg, Julian C. Cooper, Gillian Snyder, Michelle H. Lore, Dustin L. Yocum, Neal J. Cohen, Jan E. Novakofski, Melanie J. Loots, Randy L. Ballard, Mark Band, Kayla M. Banks, Joseph D. Barnes, Iuliana Bentea, Jessica Black, Jeremy Busch, Abigail Conte, Madison Conte, Michael Curry, Jennifer Eardley, April Edwards, Therese Eggett, Judes Fleurimont, Delaney Foster, Bruce W. Fouke, Nicholas Gallagher, Nicole Gastala, Scott A. Genung, Declan Glueck, Brittani Gray, Andrew Greta, Robert M. Healy, Ashley Hetrick, Arianna A. Holterman, Nahed Ismail, Ian Jasenof, Patrick Kelly, Aaron Kielbasa, Teresa Kiesel, Lorenzo M. Kindle, Rhonda L. Lipking, Yukari C. Manabe, Jade ́ Mayes, Reubin McGuffin, Kenton G. McHenry, Agha Mirza, Jada Moseley, Heba H. Mostafa, Melody Mumford, Kathleen Munoz, Arika D. Murray, Moira Nolan, Nil A. Parikh, Andrew Pekosz, Janna Pflugmacher, Janise M. Phillips, Collin Pitts, Mark C. Potter, James Quisenberry, Janelle Rear, Matthew L. Robinson, Edith Rosillo, Leslie N. Rye, MaryEllen Sherwood, Anna Simon, Jamie M. Singson, Carly Skadden, Tina H. Skelton, Charlie Smith, Mary Stech, Ryan Thomas, Matthew A. Tomaszewski, Erika A. Tyburski, Scott Vanwingerden, Evette Vlach, Ronald S. Watkins, Karriem Watson, Karen C. White, Timothy L. Killeen, Robert J. Jones, Andreas C. Cangellaris, Susan A. Martinis, Awais Vaid, Christopher B. Brooke, Joseph T. Walsh, Ahmed Elbanna, William C. Sullivan, Rebecca L. Smith, Nigel Goldenfeld, Timothy M. Fan, Paul J. Hergenrother and Martin D. Burke, Mitigation of SARS-CoV-2 transmission at a large public university. Nature Communications, 13 (1):3207, 2022/06/09 2022. https://doi.org/10.1038/s41467-022-30833-3
2. Liz Ahlberg Touchstone, SHIELD program a model for effective pandemic management, data show, in University of Illinois – News, June 9, 2022. https://news.illinois.edu/view/6367/148267842

Cheap Flexible Processors – Finally

We’ve been talking about wearable computers for a long time.  A. Long. Time [2].   But computer hardware has remained, well, hard.  I mean, it’s made out of rock, fer goodness sake!

Thus summer the clever elves down the street at Electrical and Computer Engineering report a demonstration of actually useful flexible computers [3].  These tiny 4-bit and 8-bit processors (welcome to 1970!) are fabricated on plastic rather than silicon, which makes them flexible.

Now, electronics have been produced on plastic, but only in small numbers.  The Illinois design was able to fabricate hundreds, with a high enough yield to promise a price of a penny or less.  (The research paper has a section reporting “Yield Analysis”—this is the real deal.) This would make them usable in all kinds of disposable and wearable applications, e.g., “smart” bandages, food packaging, who know what?

This breakthrough comes from the simplified design of the chip logic [1].  Rather than trying to produce contemporary 32-bit or 64-bit processors, with complex pipelines and functions, these processors are throwbacks to the good old days when the world was young and gates were expensive and difficult, so we made the most of a few components.

Super cool.

But this is from my old Alma Mater, so there’s much, much more.

For one thing, the processors (hundreds of them!) were tested with seven mini-applications, because these chips are actually programmable.  Obviously, they wrote their own assembler for these one of a kind chips, because this is Illinois and we know how to write software tools.  This made it possible to do some actual testing and benchmarking.  (There is a section, “Comparison to Existing Flexible Processors”!) 

But wait!  There’s more!

They also explored optimization of the design.  They built design tools (naturally!) make it possible to optimize the chip design for different problems.  Essentially working backwards from the code to build optimal chips to run it.  Awesome!

Very well done, all!

---

1. Nathaniel Bleier, Calvin Lee, Francisco Rodriguez, Antony Sou, Scott White, and Rakesh Kumar, FlexiCores: low footprint, high yield, field reprogrammable flexible microprocessors, in Proceedings of the 49th Annual International Symposium on Computer Architecture. 2022, Association for Computing Machinery: New York, New York. p. 831–846. https://doi.org/10.1145/3470496.3527410

2. Samuel K. Moore, The First High-Yield, Sub-Penny Plastic Processor, in IEEE Spectrum – Semiconductors, June 14, 2022. https://spectrum.ieee.org/plastic-microprocessor

1. Nathaniel Bleier, Calvin Lee, Francisco Rodriguez, Antony Sou, Scott White, and Rakesh Kumar, FlexiCores: low footprint, high yield, field reprogrammable flexible microprocessors, in Proceedings of the 49th Annual International Symposium on Computer Architecture. 2022, Association for Computing Machinery: New York, New York. p. 831–846. https://doi.org/10.1145/3470496.3527410
2. K. Van Laerhoven and O. Cakmakci. What shall we teach our pants? In Digest of Papers. Fourth International Symposium on Wearable Computers, 2000, 77-83. https://ieeexplore.ieee.org/document/888468
3. Samuel K. Moore, The First High-Yield, Sub-Penny Plastic Processor, in IEEE Spectrum – Semiconductors, June 14, 2022. https://spectrum.ieee.org/plastic-microprocessor

Successful Volcano Prediction

An absurdly successful prediction! [2]

Predicting volcano eruptions is hard.  It’s a freakin’ mountain, man!  It does what it wants, when it wants!

But sensors are getting better and cheaper, and computer models are improving. 

Down the street at my Alma Mater researchers are taking a victory lap  [1].

 In 2018, Professor Patricia M. Gregg and colleagues ported their prediction model to the local supercomputer (Blue Waters at the time), and used data from the active and well-studied Sierra Negra Volcano in the Galapagos as a test.  The forecast predicted an eruption beginning between 25 June 2018 and 5 July 2018, as reported at a workshop in December 2018.

Five months later, on June 26, Sierra Negra erupted. 

(Good volcano!  Who’s a good volcano! What good volcano you are! )

They analysis uses an ensemble of 240 models, fed Synthetic Aperture Radar data from ESA’s Sentinel-1 satellite.  This data is particularly useful for measuring the altitude of the surface, and multiple measurements reveals uplift and subsidence.  These data were analyzed based on records of earlier events.

The model flagged the June 25 time, though the model wasn’t clear whether it would be an earthquake or an eruption.  It predicted the geological failure that would produce something big.

“It is unclear whether the through-going failure flagged by the EnKF forecast was forecasting the potential of the 26 June earthquake or the eruption. We posit that the more important outcome is the success of the EnKF to quantify deformation, stress, and failure as indicators to track the evolution of the system.” (

([1], p. 6)

Cool! 

It is remarkable to me that the orbital sensors actually have enough resolution to detect the evolution of the system this precisely.  The satellite passes over every 6 days, and has limited spatial resolution.  Even the limited coverage was more data than could be processed in real time by the model.   Still, the model predicted when the rock would collapse, releasing the chamber underground.

I’ll note that the report discusses the tricky issues of interpreting the ensemble of models, as well as reporting on “hindcasts”, analyses of the data after the prediction.  It’s all pretty neat, though I don’t necessarily grok the details.

The researchers indicate that this cool result was possible only because of the availability of oodles of computing power.  The model was originally developed on laptops, but real time prediction needs a lot more oomph than that.

The researchers are thinking of adding in machine learning to improve the forecasting.  If it works this well without ML, think what may be possible with a big enough training set?

---

1. Patricia M. Gregg, Yan Zhan, Falk Amelung, Dennis Geist, Patricia Mothes, Seid Koric, and Zhang Yunjun, Forecasting mechanical failure and the 26 June 2018 eruption of Sierra Negra Volcano, Galápagos, Ecuador. Science Advances, 8 (22):eabm4261,  https://doi.org/10.1126/sciadv.abm4261
2. Lois Yoksoulian, Great timing, supercomputer upgrade lead to successful forecast of volcanic eruption, in Illinois News Bureau – Research News, June 3, 2022. https://news.illinois.edu/view/6367/913924091

More on the Costs of Machine Learning

Machine Learning has been the flavor of the month for a couple of decades now.  A big reason for this has been scaling.  Moore’s Law and all that.  Basically, whatever you are doing with Machine Learning, you can do way more of it now than you could ten or twenty years ago.  And that has meant that lots and lots of problems have been tackled with amazing results.  (Not always entirely correct, but amazing nonetheless.)

But scaling isn’t free, and this is increasingly creating issues.

For one thing, all that magic is using resources.  It is becoming important to account for the inputs (the costs), not just the outputs (the benefits).  Just making things bigger, using more data, and running more trials, isn’t necessarily going to produce proportional benefits.  It may not even be sustainable. 

This accounting isn’t trivial.

This arms race has a second effect.  In many cases, advancing research requires using the largest machine learning available.  While anyone can do ML on their lap top, the truly groundbreaking magic requires gigantic systems, available only to a few researchers.  And, increasingly, academic researchers do not have such access.

Now, academia is always begging for resources, and frequently has to do deals to get hands on tools.  But in the case of large scale ML, there simply aren’t systems available, even if researchers had the money, or are ready to grovel collaborate.  Which is a huge problem, because academia is where the stuff comes from, and if we can do it now, no one will have it ten years from now.

This fall Vincent J. Hellendoorn and Anand Ashok Sawant discuss yet another case: machine learning applied to software engineering [1].

This topic can mean a lot of things, but one rapidly advancing area is the automatic generation of code based on vast datasets of example code.  The basic idea is that source code is text, so ML techniques can learn to generate code that is good as humans write.  If nothing else, this approach can generate souped up “autocomplete” facilities. And, at best, it can save huge amounts of time, generating the 80% of code that is purely routine, leaving programmers more time to worry about the hard bits.

H & A report that these efforts use gigantic datasets of example code, and run gazillions of cycles to train the models.  The monetary cost is surely millions per model.  And you probably need thousands of runs to discover the optimal parameters.

“This may be a small price to pay for Facebook […] but this exploding trend in cost to achieve the state of the art has left the ability to train and test such models limited to a select few large technology companies—and way beyond the resources of virtually all academic labs.”

([1], p. 31)

One prescription is, well, money.  Basically, there is a need for public, shared computational resources.  For ML, this means fairly large clusters with lots of GPUs.

In that light, I note that as the Blue Waters Supercomputer shut down, the National Center for Supercomputing Applications is replacing it with Delta, a big GPU cluster.  This system will not be dedicated solely to ML, but it looks like it could be just the thing for larger academic ML studies.  I’ll add that NCSA has a long history of expert user support, including considerable experience with large arrays of GPUs.  (NCSA calls this “useability”, i.e., making the system useful to as many researchers as possible.)

So, it looks like the powers that be in the US are responding to the needs of academic researchers, at least this much.  And my local supercomputing center is still in the mix.

---

1. Vincent J. Hellendoorn and Anand Ashok Sawant, The growing cost of deep learning for source code. Communications of the ACM, 65 (1):31–33,  2021. https://dl.acm.org/doi/10.1145/3501261