|
10th Anniversary/Millennial Survey
First, let's look at how SC&G has changed
over the past 3 decades. Questions (1) and (2) are worded in terms of
systems. If you prefer, you can instead address successful
ideas or methodologies for SC&G
.
- What system or feature of a system marked an
especially important advance, for its time, in our practice of
SC&G ? How did that idea evolve in more
recent systems?
Andreas Buja, AT&T Labs -- Research
Raster graphics. Everybody takes it for granted now. You kids don't
remember, but before 1980, there were only expensive vector graphics
machines such as the legendary Evans and Sutherland. Realizing the
importance of the move to raster graphics, Werner Stuetzle, John
McDonald and Jerry Friedman at Stanford put together their own
hardware for a workstation with a raster display in 1980/1 and with
Andy Bechtolsheim's SUN board number 1, I believe. This effort was
called the ORION project. Everybody knows about PRIM-9, but ORION was
of even broader importance by spearheading workstations with color
raster displays for statistics.
Di Cook, Iowa State University
The evolving of interactive "direct manipulation" and dynamic
graphical methods. I was in Canberra at the Australian National
University in the mid-1980s working as a research assistant to the
Statistics Dept. I had just discovered graphics terminals, tektronix's
especially, and was making lovely colorful contour plots of likelihood
functions. I believed they were so beautiful, and was a little
disappointed to see some of the statisticians in the group prefer the
horrible printer plots of contours represented by letters. Then Bill
Cleveland visited, and gave a talk about linked brushing in
scatterplot matrices. Several of the statisticians got quite
excited. But I thought them to be ugly, and didn't see the
significance until coming to the USA in 1988. Then I discovered Sun
workstations, and the X Window System, and met Javier Cabrera who
encouraged me get back into statistics. Working on a 3D graphics
program, I was introduced to Andreas Buja, Debby Swayne, later Rick
Becker and Dan Carr and through these people suddenly realized why
statisticians need linked brushing, and learned more about the
research into other methods like this. The evolution of the X Window
System allowed us to develop these tools into a package called XGobi,
whichprovided wide accessibility of high-level interaction and dynamic
graphics to every unix workstation. Unfortunately glossy graphics has
not disappeared, and is probably more prevalent than the
high-interaction graphics, especially in the visualization community
and maybe data mining. I think this causes some confusion about the
importance of graphics for exploratory analysis. It is important to
note that points and line plots, despite their "ugliness," have far
greater universality for data analysis, for representing arbitrary
dimensions. Glossy 3D graphics are great for rendering physical
phenomenon but don't extend well to arbitrary dimensions, as is needed
for working with data that are measured on any number of
variables, not just 3. I think this is not fully realized
yet.
- Thinking about ideas that have been particularly successful leads
one naturally to consider "features" that have gone away. What would
be an idea or characteristic in past systems that is
notable for disappearing (whether or not, of course, you are
glad that it did)?
Andreas Buja, AT&T Labs -- Research
The integrated programming environments of the Lisp machines produced
by Symbolics and TI. It is tragic how a superior technology
disappeared. Neither Unix nor DOS/Windows can measure up with the
seamless integration of compiler, loader, linker, interpreter,
debugger, editor, window system, networking, all formulated in a
single language, namely, a highly developed Lisp dialect with OO
features, that was both assembler at the bottom and applications
programming language at the top, and that ran interpreted as well as
compiled. C++ by comparison looks like an awkward misunderstanding of
OO concepts. At the higher level we're lucky to have had the S,
Lisp-Stat and R developments which really gave practicing
statisticians the interpreted languages they needed. Had Lisp
machines existed a little longer with wider distribution in
statistics, we would have felt their demise much more.
Another example of dead technology are expert systems, but in this
case I won't shed a tear. I was always half-hearted about them,
although I was involved on the periphery. Now I see that I was just
part of a greater lack of enthusiasm which undid the fad. Replacing
reasoning humans with reasoning computers was a bad idea. No one
wants it. Extending humans' cognitive abilities is a good idea,
however, which is why visualization succeeds.
- Work in SC&G is
obviously limited by available programming languages, operating
systems and computer hardware. What was the tallest technological
hurdle you've (personally) cleared? What were you trying to do
and how did it come out?
Andreas Buja, AT&T Labs -- Research
I don't clear hurdles. I wait till things offer themselves in a
reasonable work environment. Others have more pioneering spirit.
- Looking back over the past 10-20 years, what do you see as the
single most important change in "technology" (hardware,
software or something in between) supporting SC&G
?
Andreas Buja, AT&T Labs -- Research
Platform independence. We shouldn't have to worry anymore about
Windows, Mac, Linux, Unix,... Everything should run everywhere. Java
and CORBA (with its OMEGA tentacle in statistics) are certainly great
promises.
Di Cook, Iowa State University
I think Sun workstations. They had a glamour. We lusted over them in
Australia, only seeing them for demonstrations in Australia in the
mid-80s. They made graphics more "interactive," I think. The old
tectronix terminals were specialist graphics devices, so one had to
quit all other work and go to them to do the graphics and back to the
old vt100 to do the numerical work. Also the X Window System
which beat out NeWS, the postscript window system. It made it
possible to run applications remotely, and standardize the graphics on
unix machines. Interestingly these are "open source" software. I think
for the future to provide better collaboration amongst statisticians
we need open source statistical software for both numbers and
graphics. The advances that have been made in the computing community
to unify machines may have stemmed from the open source
mentality. Thats not true with respect to Microsoft where market
dominance has led to standardization, but what a crummy standard. R
is a great example of the collaboration that can be achieved between
statisticians when software is provided as open source.
And now an opportunity to take stock of SC&G
, its present and future.
- What do you see as the most striking change taking place
in the field of SC&G ? What is the most
important development in the theory and/or practice of
SC&G ? Are these changes being driven by
new technologies or new statistical applications, or
both?
Andreas Buja, AT&T Labs -- Research
Bayesian computing. I'm not one of them, but they are sure taking
over. It's just an outgrowth of crude CPU speed. It seems
simulation, whether MC or MCMC, is now actually replacing analytics to
some extent. The part of asymptotic theory that has to do with
asymptotic distributions intended for testing is becoming obsolete.
In effect, asymptotics is reduced to a kind of "metaphysics" of
statistical methods, where proofs of consistency and asymptotic
efficiency have utterly no impact on practice. An exception has been
bootstrap, where asymptotic theory showed why bootstrap doesn't work
for certain discontinuous statistical functionals.
- What do you predict will be the single most important
change in "technology" (hardware, software or something in
between) supporting SC&G and what will its
long-term impact be?
Andreas Buja, AT&T Labs -- Research
Internet. Dissemination, communication, collaboration thru web
technology.
- What do you see as the most important
contribution SC&G
can make to the statistics profession as we enter the new
millennium?
Andreas Buja, AT&T Labs -- Research
Simulation and visualization.
Bayesians have caught on to the power of simulation, but frequentist
methods can also be simulated, often more easily than Bayesian
methods. There's potential for MC especially in the area of exact
tests, which exist, for example, for regression coefficients and
balanced random effects.
Data visualization should not really be a subdiscipline of statistics
and data analysis. The goal of data viz will be achieved when it
disappears as a specialty. It should really be part of all good data
analysis, and statistical methods should be more tuned to the
production of qualitatively rich intermediate aggregates suitable for
visual display, as opposed to single numbers, be they estimates or
p-values. Actually, the situation is not so bad. We have come along
way with research into "intermediate aggregates" such as density
estimation, smoothing, and other function fitting. The ultimate goal
of densities and smooths is often visual display, not prediction, and
they therefore feed data visualization. What's missing is theory that
allows one to assess observed features in densities and smooths.
Marron is one of the few who has drawn the full consequences and
reoriented his research accordingly by dumping MSE-oriented theory and
developing inference for observed features.
Next, we reflect on the state of our profession.
- SC&G , like any field, is really about
people: the people who contribute to it and the way they relate to the
wider community it serves. What do you see as the most important fact
about the personal environment for contributors to SC&G ?
Andreas Buja, AT&T Labs -- Research
Luck is always the most important fact: luck with finding congenial
and complementary co-workers. I feel I have been lucky due to my
association with Debby Swayne and Di Cook, and previously to the
inspiration from the ORION project of Werner Stuetzle, John McDonald
and Jerry Friedman. Something I'm missing a little is a focal
personality such as JWT, but that's just a personal sentiment. The
community at large can probably progress reasonably well without
towering figures.
- What can be done to ensure the vitality of SC&G as a profession?
Andreas Buja, AT&T Labs -- Research
Increase computational literacy among young researchers. The current
situation is probably not so bad. Young PhDs seem pretty unafraid of
computers. Except: I'm a little disappointed how few people have
tried an interactive data visualization system (not necessarily mine).
Once they get going with a little advice, they get quickly
enthusiastic, but that little investment is not happening in grad
schools. Of course that's due to the absence of teachers with
expertise. Building such expertise should really be Dan Carr's and my
crusade, but unfortunately we're only mediocre proselityzers.
In designing this questionnaire, we have likely left things
out. Please use the space below to add other thoughts, memories, or
predictions that we didn't cover with one of the questions above.
- Other comments?
|