Making Sense of IoT Data with Machine Learning
A big promise of the Internet of Things (IoT) is that by analyzing millions of new sources of data from embedded, networked devices our experience of the world becomes better and more efficient. The environment automatically predicts our behavior and adjusts to it, anticipating problems and intercepting them before they occur.
The notion is seductive and almost magical: an automatic espresso machine that starts a fresh latte as you’re thinking it’s a good time for coffee; office lights that dim when it’s sunny and electricity is expensive; a taco truck that arrives just as the crowd in the park is getting peckish. Exciting in theory, this promise is rather unspecific in the details. Exactly how will our experience of the world, our ability to use all the collected data, become more efficient and more pleasurable? However, we don’t have good examples for designing user experiences of predictive analytics.
At RE.WORK Connect Summit, in San Francisco on 12-13 November, Mike Kuniavsky, Principal Scientist of Innovation Services at PARC, will present on the importance of predictive behavior to consumer IoT products and services, exploring UX challenges to creating such behavioral systems, and suggest patterns for addressing those challenges. We caught up with Mike ahead of the summit next month to hear more about his work at PARC.
Tell us about your work at PARC and the Innovation Services Group you operate within.
Broadly, the mission of PARC’s Innovation Services Group is to reduce the risk of adopting cutting edge technologies using a combination of ethnographic research, user experience design, and innovation strategy. Essentially we’re PARC’s consulting arm, which matches PARC’s deep bench of novel technologies with our customers’ specific business needs, using user-centered methods.
Because of our deep hardware, networking and machine learning background, these days we mostly design Internet of Things products and services for Fortune 100 commercial clients (of which Xerox is one, of course, though they’re not our biggest client). As the senior UX designer/consultant on the team, this means that my work focuses on both broad service design ideas, such as creating consistent experiences that span multiple devices and apps, and on specific design challenges that lie at the heart of such offerings, such as how to design a wearable device (or a device for the home, or car, or long-haul trucking cab, etc.) that uses a novel sensing technology (a PARC specialty) that’s never existed before.
How can we make sense of IoT data with machine learning?
Humans are good pattern matchers at certain things, but we’re not built to collect and make sense of huge amounts of data or to articulate our needs as complex systems of mutually interdependent components. Computers are great at these things. They can make statistical models from many data sources across space and time and then maximize the probability of a desired outcome. The IoT produces rivers of data that are constantly shifting, with new patterns that dynamically emerge and old ones that dissipate. It’s incredibly difficult for someone to make sense of that flow, but that information often can be critical to the livelihood of individuals. Broadly speaking, machine learning is the umbrella term for algorithms that automatically (or with some human help) identify patterns in these data rivers and determine which device behaviors tend to create the most desirable outcomes. Models that learn and continuously adapt from the outcome of thousands of situations across many people and long periods can compensate for a much wider variety of situations, in a more nuanced way, than just about any individual will ever be able to. This helps such systems be great tools and assistants–and, of course, there are more nefarious applications to machine learning and artificial intelligence (the larger discipline of which machine learning is a part), but as user-centered developers, we’re careful at PARC to minimize the known negative effects of the systems we build and maximize their utility to consumers.
In your talk you will discuss UX of predictive analytics. What are the challenges to this user experience design?
Because predictive machine learning systems are statistical in nature and change dynamically based on large sets of data, they’re quite opaque, often even to the developers of the algorithms making the models. When there are no obvious dials or configuration levers with which to adjust their behavior, and the entire interaction paradigm of working with a system that changes its behavior autonomously, that learns and may act differently tomorrow than it did today, the user experience can be very confusing and frustrating.
We work a fair bit with machine learning systems, and the challenges run the gamut of UX: on the service design level, do we treat a machine learning system as a core actor, one that makes decisions on people’s behalf, or an assistant that recommends courses of action? The answer depends on many factors. On the interaction design level, do we ask people to train a system, as they would a puppy, for the first month, with the expectation that initially it’ll act unpredictably most of the time, even though the results will be much better after that month than if we used a generic model that works out of the box?
How can we learn to embed consumer behaviour in design foundations?
Design is the practice of aligning the experience of a product with the business needs of the organization creating it. It’s about identifying how the extra effort necessary to use a new device or service will be worth it to an intended audience (because, after all, managing time is the ultimate zero-sum game, and if we’re asking people to use our product, we’re simultaneously asking them to NOT do something else, so we have it make it worthwhile). Devices that respond to our behavior can be great assistants if they match the mental model of what we intend to do, or huge annoyances if they require constant management. By embedding dynamic behavior, and a form of intelligence, in our devices, we open up huge potential for creating beneficial experiences, and potentially incredible frustration if things don’t go well. Right now we’re at the earliest stages of understanding how to do that, so there are few answers, but I believe we’re starting to see the outline of at least the early challenges.
What sector or industry has the potential to have the most significant disruption by IoT?
It’s tough to say (if I knew for sure, I’d be investing all my money instead of working at a research lab), but I believe that there are a handful of obvious places where we’re starting to see it: the integration of IT systems at hospitals is creating some very IoT-heavy environments where the information affects people’s health directly; manufacturing and the operation of other heavy machinery (such as passenger cars) is replacing people who do highly repetitive actions with robots that are much more intelligent than simple movable arms (the Baxter robot, etc.); adding small amounts of predictive intelligence is changing the way that everyday consumables, such as diapers or bottled water, are being consumed. Xerox printers have been able to predict when toner is going to run out for many years, and then automatically reorder it so it arrives just in time. I suspect that’s coming to everything from potato chips to gasoline.
Mike Kuniavsky will be speaking at RE.WORK Connect Summit, in San Francisco on 12-13 November. Use the discount code IOTDO for 20% off, visit the event website and book now!
For further discussions on Internet of Things, Wearables, Connected Devices & more, join our group here.