16 July 2015

by S. Alspaugh


Last night Three weeks ago1 was the HCIC opening keynote by Stuart Card. As the theme of HCIC 2015 was "theory", his talk was appropriately titled "Theories with a Purpose".2

HCIC stands for Human-Computer Interaction Consortium. It is an intimate, multi-day, annual conference for the HCI community. Currently in its 26th year, HCIC was created by researchers who wanted a venue that emphasized conversation over presentation. For many years, HCIC was held in Breckenridge, Colorado, also known for its high-quality skiing. If you want more of the inside scoop, you'll have to ask Stu Card.

Stu Card is one of the founding figures of HCI. From his bio:

Stuart Card may have pursued the first direct program at CMU in Human-Computer Interaction—a program involving Ph.D. qualifiers in both psychology and computer science. ... Card, Moran, and Newell wrote the first book to use human-computer interaction in its title, The Psychology of Human-Computer Interaction.

Card has a marvelous sense of humor and a gift for storytelling that is enhanced by the fact that, thanks to his long history in the field, he is loaded with humorous anecdotes that include many early research icons.

Pretext

Card began his talk by describing some playful banter between himself and Judy Olson.

Judy is a highly-respected senior HCI researcher and professor in the Informatics Department at the University of California, Irvine, and who also happened to be in attendance. She recently published a book entitled Ways of Knowing in HCI, which describes research methods in HCI, with each chapter detailing a different approach.

Card had complained to Judy that the chapters of this book were ordered in a seemingly random way. He also noted, as is common for academic books, that it was rather expensive. So he proposed a wager. If, in his talk, he could develop a theory from first principles for better structuring the Table of Contents, then Judy would give him the book for free. Judy would be the judge of whether or not he had succeeded.

On developing a new academic fields

There is a process to defining a new field. At first, there a chicken-and-egg problem, as textbooks must be written so that courses can be taught, but courses must have been taken in order to know enough to write a textbook. Textbooks are needed for uniformity among courses at different universities, otherwise it isn't a field. Lastly, journals need to be created so that researchers in that field can get tenure. This was once the case with HCI too.

Here is a good trick for defining a new field: Define the core of what the field is about, then define the distance of various topics from the core, then let what is considered topical "fade out" as you get further from the core. This is as opposed to defining what is in and what is out. The former approach makes is easier to absorb new research topics into the field.

The above discussion brings to mind current efforts on Berkeley's campus to develop a new data science curriculum.

On a theory of artifacts

Card described the basic principles of operation of the old-fashioned optical mouse. He segued this into a discussion of Herbert Simon's ant. Consider the following line:

What does it describe? It appears to be a fairly complicated line that could be difficult to model and simulate. In fact, the line it describes the motion of a ant on a beach, whose purpose is to traverse the beach, but whose path must twist and turn around the rocks on the beach.

The algorithm it uses to do so is simple: walk down the beach while avoiding the rocks. Note that the complexity is not in the behavior of the ant. It's in the beach. The apparently complexity of the ant's behavior is is actually just a reflection of its environment.3

All artifacts have: a purpose (walk down the beach), an inner environment (walk home while avoiding rocks), and an outer environment (the beach).

Card then described the inner operation of the laser mouse: entirely different inner environment, but same purpose.

"This is what I love about computer science and computational fields. You can do the same work over every twenty years and it's like a new area. Because everything is faster."

Stuart Card

On dominant design

There are four phases of design research:

  1. Antecedent research: This phase contains the early days of experimentation.
  2. Pioneering architecture: This represents first attempt at breaking systems down into independentally optimizable subsystems, which people can optimize into even more subsystems.
  3. Pioneering product: Companies productize different designs with different architectures. One company wins.
  4. Dominant design: The design has become locked in and cannot be changed anymore.4

According to Card, we are in a new phase of antecedent research for mobile and immersive interaction technology.

Decomposability into subsystems is very important for being able to optimize a design. The champions of such decomposition are aircraft designers.

Card then described Herbert Simon's hierarchy:

social
social behavior with a time scale of days to months
rational
problem solving with a time scale of minutes to hours;5 example theories: information foraging and sensemaking
cognitive
immediate behavior with a time scale of seconds; example theories: Goals, Operators, Methods, and Selection rules (GOMS), window, and working set
biological
automatic behavior with a time scale: milliseconds; examples: Fitts' law

These are nearly decomposable. In each case we can apply theories to:

  • Interpret data (understanding and notation)
  • Generate new artifacts
  • Predict performance
  • Unify with other domains

On theory versus experimentation

Theory trumps experimental results in demonstrating that option A is better than option B. Card illustrated this with a story: He was tasked with presenting to a room of fifty of the most hostile engineers imaginable about the superiority of the mouse as an input device. The engineers didn't want to design a mouse for their system, they wanted to bolt on a keyboard and then bolt on (from what I understand) a sort of trackball. Card showed them many strong experimental results6 demonstrating the superiority of the mouse for interacting with computers over their preferred systems, but the hostile engineers thought of a million reasons why the experiments might be wrong. It wasn't until he explained the theory, and why the theory shows that the mouse has to be the right answer, that the engineers went along.

The theory (Fitts' Law) says the fastest input device would use the fingers, because these are the fastest instrument of implementing desired plans of action as they are conceived of in the brain. Fingers have a throughput of 42 bits per second. So what if you got even closer to the brain— via a head mouse, perhaps? Back when he was at Xerox PARC, Card had a colleague who was a great champion of the head mouse. Card thought this was not a good idea and that it was unnecessary to try, because the theory said it wouldn't be better than the mouse. This is because the arm and wrist that moves the mouse has a throughput of ten bits per second, but the head mouse is actually moved by the neck muscles, which only have a throughput of five bits per second. But humorously, said colleague persisted, and ended up walking around for a week with a sore neck. The head mouse idea was scrapped; the theory was correct.

"When you do theory, it can actually be fun to be wrong if you can be really definite."

Stuart Card

One theory can tell you the best option out of hundreds of options. This can save you from running hundreds of "expensive" experiments with ten whole participants and controls for learning effects.

"Experiments tell you the answer to the question you asked, but that is not what you want to know. You want to know the answer to the question you should have asked."

Stuart Card

The big reveal

In the end, Card linked contemporary research7 back to early work on the Model Human Processor and to the GOMS framework, and related these foundations to contemporary work on sensemaking.8

Somewhat tounge-in-cheek, he illustrated this with the following hand-drawn diagram:9

Returning full circle to the original wager, Card finished up by using these concepts he to group chapters from Ways of Knowing in HCI, into subsections:

In the end, Judy graciously acknowledged that he had a good overview and an interesting perspective. Her definitive verdict was that he did not have to pay for the book.

In the questions after the talk, someone asked, in Card's opinion, what is missing from the methods in the book? His answer is that learning i.e. probabilistic methods or Bayesian methods need to be included in the future, given that there seems to be a lot of evidence that that stuff "just works".

Acknowledgments

Thanks to Stuart Card for his lovely talk and input on this post, as well as for sending me his slides. Many thanks to Elizabeth Churchill for kindly helping me improve this post.


Footnotes

  1. ^ I fell a bit behind in posting this due to a paper deadline.
  2. ^ I recently discovered a video of an identically-titled talk, that Stu gave last November at CMU's Human-Computer Interaction Institute's 20th anniversary celebration. Though the high-level topic appears to be the same, the details of that talk seem to differ substantially from the one I describe here.
  3. ^ This analogy is courtesy of Herbert Simon. I don't know where it was originally published, but in this supposed interview with Simon from 1997, he elaborates upon the idea. It may have originally appeared in his book Administrative Behavior.
  4. ^ My interpretation of this is that here he is talking about conventions for the design of objects, my views on which are entirely informed by The Design of Everyday Things.
  5. ^ Where my work sits.
  6. ^ Many of the early experiments on the design of the mouse were done by Bill Moggridge, co-founder of famed design firm IDEO. IDEO designed the first production mouse for Apple. Moggridge wrote a book, Designing Interactions, about the history of interaction design, which contains interviews with a number of famous figures including Douglas Engelbart, Larry Tesler, and Stu Card, among others.
  7. ^ Contemporary work that got a shout out in the talk included Skinput, the Proteus Ingestible Networked Pill, and ShapeWriter.
  8. ^ His book with pioneers Allen Newell and Tom Moran, The Psychology of Human-Computer Interaction, details the beginning of GOMS, which became significantly more elaborate as others extended it. One of the earliest experimental successes in HCI were the NYNEX experiments, in which researchers used GOMS to predict how efficient telephone operators at NYNEX (the local telephone carrier for much of New England and New York state) would be by using expensive new workstations, versus their current method.
  9. ^ As far as I can tell, it seems to be a habit for papers about sensemaking and related topics to contain a diagram illustrating the sensemaking cycle. I vaguely remember drawing one such diagram at one point to explain some idea I had to my advisor, Marti Hearst. She waved it away saying something to the effect of, "oh, that's just the same diagram that everyone draws."