Pay Attention to the Man Behind the Curtain
By Henry Lieberman
MIT Computer Science and Artificial Intelligence Lab (CSAIL)
Machine learning often seems like magic — and it is.
Machine learning is often applied to prediction problems. In the past,
those problems would have been approached by careful prior analysis of
the problem domain, trying to analyze the underlying processes, and
making rules and representations to perform inference.
With modern machine learning techniques, once an algorithm,
parameters, and data sets have established, for a wide range of
problems, you can just push a button and get an answer. That answer
turns out to be right, a certain percentage of time. It feels like
pulling a rabbit out of a hat.
But like the Sorcerer’s Apprentice, you have to be a bit careful what
you wish for. As with stage magic, the trick may have some limits.
The magician won’t explain to you exactly how the
trick works. Often, even machine learning researchers can’t supply
plausible explanations for why their techniques succeed.
The trick may be dependent on certain aspects of the situation which
you may not have noticed, but are crucial for the trick to
succeed. You may not have been paying attention to the alignment
between the bottom of the hat and the table it rests on, but if you
move the hat, the trick might not work. It may not work for problems
that seem similar but really aren’t. The magician may be able to cut
a lady in half, but that doesn’t mean they could cut a tiger in half.
If you want to learn how to become a good magician, you can’t just sit
back and be wowed by the trick. You have to dig into the details of
how it happens, and what it can and can’t do. Successful machine
learning takes some knowledge, and a lot of experimentation.
And while magicians do mind-reading tricks, machine learning can’t actually
read your mind. So there needs to be some communication between the
machine, and people who serve as its trainers, implementors, and users.
Recently, a movement has arisen to focus on the human aspects of the
machine learning process: Human-Centered Machine Learning.
This TiiS special issue provides an introduction to this topic, ten articles
that envision a new partnership between people and machine learning
algorithms. They show how judicious use of human interaction can lead
to successul applications of machine learning in a wide variety of
The first article is a survey of interaction design for machine
learning, by Dudley and Kristensson. This is the place to start, along
with the issue introduction by Gillies and Fiebrink. If you only have
time for one article, this is it. They deliver six principles that
serve as guidelines for UI design for applications. If you’re a UI
designer tasked with machine learning interfaces, this gives you
something to post on your wall.
The rest of the articles cover a wide variety of applications, from
social science to medicine to visual and audio presentation. There
are nine of them, so there won’t be space in this post to even give a
summary of each one. See the issue introduction for that.
The breadth of topics covered give you a feel for how each approached
the design challenges of human interaction. They might be an
eye-opener, if your whole experience with machine learning is reading
about mathematical algorithms, and playing the one-upmanship game of
comparing accuracy rates on standardized datasets.
Many of them center around the machine learning practice of “labeling”
data. This is taken for granted in many machine learning
discussions. But the process of asking a person to label something
brings up a whole host of issues: labels can be ambiguous or
misleading; people may disagree about labels; new concepts may arise.
Some introduce innovative interfaces for doing the labeling
interactively, presenting the task in user-friendly ways, and
providing ways to assure that the labels meet the users’ intent.
See particularly Chen, Dumitrache, and Kim.
Also necessary is to present the output of machine learning in
user-friendly ways. Visualization uses the interpretive power of
the visual system to integrate information in the blink of an
eye. Presenting results in a manner appropriate to the application context
helps give users an understanding of the confidence they should
have in results, and how far they can be generalized. See Morrison,
Smith, and Francoise.
Not to be neglected is what happens in between the labeling (input)
stage and the prediction (output) stage. Many machine learning
techniques come with a 747-cockpit of parameters and modeling
choices. Most of the articles do treat the question of how to manage this
complexity, from the perspective of the machine learning approach they
have chosen, and for the application area they are working in. We have to
get past the temptation to just treat machine learning as a black box.
Arthur C. Clarke said, “Any suffficiently advanced technology is
indistinguishable from magic.” Magic is fun, and it’s undeniably
impressive. But it’s Human-Centered Machine Learning that wil get us
to understanding machine learning as the sufficiently advanced
technology that it is.