Big Data, Big Design

Why Designers Should Care about Artificial Intelligence

Helen Armstrong, Keetra Dean Dixon

• Publisher: Princeton Architectural Press
• Year: 2021
• ISBN:
  9781648960789
• Topic: Design
• Subtopic: Graphic Design

Big Data, Big Design provides designers with the tools they need to harness the potential of machine learning and put it to use for good through thoughtful, human-centered, intentional design. Enter the world of Machine Learning (ML) and Artificial Intelligence (AI) through a design lens in this thoughtful handbook of practical skills, technical knowledge, interviews, essays, and theory, written specifically for designers. Gain an understanding of the design opportunities and design biases that arise when using predictive algorithms. Learn how to place design principles and cultural context at the heart of AI and ML through real-life case studies and examples. This portable, accessible guide will give beginners and more advanced AI and ML users the confidence to make reasoned, thoughtful decisions when implementing ML design solutions.

CHAPTER ONE Peek Inside the Black Box

Each day we generate data—terabytes of it. How have you produced data in the last month? In the last week? In the last hour? Did you write an email? Post a photo? Text a friend? Watch a streaming video? Wear an activity tracker? Drive through a traffic camera? As we move through our lives, we leave behind a garble of unstructured data—i.e., data not organized into ordered sets like spreadsheets or tables. Scholars claim that as much as 95 percent of all data is unstructured.¹ Machine learning (ML) enables a computer to derive meaning from all this unstructured data. Even now as you read, computers sift and categorize your data trails—both unstructured and structured—plunging deeper into who you are and what makes you tick.

FIG 1. STITCH FIX ALGORITHMS TOUR. Through interactive storytelling, Stitch Fix visualizes its use of rich data to match clients with items of clothes, shoes, and accessories. The company combines algorithmic decision making with human skills—intuition, understanding context, and building relationships—to make shopping personal.

Today, computers intuit the world more like humans. When I enter a room, I don’t learn about the room via a spreadsheet. Instead, I use my senses. I analyze images, sound, space, and movement. I take this information and make decisions based on what I find. Combining sensors (accelerometers, barometers, gyroscopes, proximity sensors, heart rate monitors, iris scanners, ambient light sensors, chemical and microbial sensors, electric noses) and other input devices (cameras, microphones, touch screens) with ML turns each trail of unstructured data into a richness of organized, coveted data resources. Imagine the impact of transforming vast quantities of previously unusable data—your data—into information that can be detected, digitally stored, and analyzed. Your politics, your personality, your sexuality, your next move. This is the future that is materializing right now.²

Without the sheer quantity of this data—data that used to be lost in the digital abyss—ML could not function effectively. Why is this? ML algorithms train using examples—bundles of data. The size and range of the examples determine the subsequent accuracy rate.³ This training process also requires masses of “compute,” i.e., resources and processing power to fuel complex computation. According to journalist John Seabrook, “Innovations in chip design, network architecture, and cloud-based resources are making the total available compute ten times larger each year—as of 2018, it was three hundred thousand times larger than it was in 2012.”⁴ ML has taken off recently because both data collection and compute, along with accessible and affordable input devices and sensors, now flourish in our society.

THE ONSLAUGHT OF ALGORITHMS

But how does this mysterious ML stuff really work? Yesterday I checked my email, searched for an old high school friend online, used Waze to get across town—and tip me off to where the cops were—checked my Instagram feed, asked my Amazon Echo about the weather, and got a fraud detection alert from my credit card. Which of these activities involved ML? All of them. A prediction facilitated each of these interactions.

Put simply, ML consists of algorithms—in essence a set of task-oriented mathematical instructions—that use statistical models to analyze patterns in existing data and then make predictions based upon the results. They use data to compute likely outcomes. We can think of these algorithms as “Prediction Machines.”⁵ These algorithms might predict, for example, the buckwheat pillow that you are likely to buy or the Netflix series that you will binge next. They might predict the arrival time of an Uber or whether or not an email is spam (and whether you’ll open it). They might predict the identity of a face or even the profile that will intrigue you on Tinder.

The magic of these predictions lies in the learning. ML algorithms not only analyze historical data, they also, once trained, make predictions about new data. For example, an email platform might employ ML to detect spam. The trained algorithms will be able to sleuth whether or not an email should go straight to the junk folder, not only in the original set of training data but also in new data—new emails—that enter the system.

Again, ML algorithms need to feed on a large quantity of training data to attain a high accuracy rate of predictions. Let’s go back to humans for a moment as we consider this. In order to intuit things about the world, humans observe and interact with their environments. We noted earlier that this often occurs via unstructured data—movement, sound, images, etc. The wider the range of data that we encounter, the more complex our understanding grows. For example, imagine spending your life in one small room, say your bedroom, interacting with only one person, perhaps your dad, over the course of your life. Everything you learned about human behavior would come from that single environment and that single person. Your understanding of the world would, subsequently, be very limited. The same is true for predictive algorithms. If, for instance, you build an ML system to identify a range of objects but only supply images of cars as training data, the system will subsequently classify every object it comes across as a car—because cars are all it knows. If training data only includes a narrow slice of examples, the system will only be able to make predictions that relate to a small range of possibilities.⁶

How does ML compare to traditional programming? Traditional programming requires developers to enter explicit logic-based instructions—code—to produce behavior within a software system. In contrast, ML empowers the computer to observe and analyze behavior happening in the physical or digital world and then produce code to explain it.⁷ We can say, “Hey, computer, make a prediction based upon these examples and then produce code that can apply that same prediction to future data.”⁸ This approach allows ML to take on predictions that are too complicated to address through line after line of logic-based instructions, like identifying a human face or determining the meaning of your last query to Alexa.

ADD A LITTLE PREDICTION AND SEE WHAT HAPPENS

When combined with intriguing datasets and powerful visions, ML algorithms can be quite transformative. Artists/ activists Mimi Ọnụọha and Diana J. Nucera (a.k.a. Mother Cyborg) in their zine, A People’s Guide to AI, compare this kind of technology to salt—less interesting on its own but once added to food, “It can transform the meal.”⁹ Imagine that while watching a baseball game a prediction pops up that your favorite player is about to strike out and strand two batters on second and third bases. How would this kind of predictive power change the way you view the game or how players strategize their next move? Salt works as a metaphor for predictive algorithms on another level as well. Salt flavors many meals because it is cheap and widely available. As the cost of prediction technology drops, we will use it more and, as we increase use, its impact will compound.

“Abundant, cheap predictions are going to change the material that you’re designing with, and you better get used to understanding how to work with it.” —Tony Chu, Facebook

Ajay Agrawal, Joshua Gan, and Avi Goldfarb point to this commodification of prediction in their book Prediction Machines. They look to artificial lights as a precedent. In the early 1800s, artificial light cost four hundred times more than it does today.¹⁰ When the cost of artificial light plummeted, human work and family habits, as well as architecture and urban planning, changed dramatically. We could suddenly build rooms without windows and create large structures within which we could live and work both day and night. Hello, night shift.¹¹ We saw a similar phenomenon in the 1990s, when the internet lowered the “cost of distribution, communication, and search.”¹² Whole industries, such as those for pagers, encyclopedias, and answering machines, disappeared or were revamped—video stores to Prime Video, travel agents to Google Travel—to take advantage of these new cheap capabilities. Consider the ramifications of the internet on music, catalog shopping, money transfer, postal services, archiving, etc. As the cost of utilizing ML algorithms drops, we can begin to imagine the impact of powering up prediction on many aspects of our lives.

DESIGNERS NEED TO JUMP INTO MACHINE LEARNING

ML needs designers. Our human-centered methods articulate human needs and desires in relation to the larger society. Too of...