Saturday, April 13, 2019

Out Of The Sandbox: AI For The Real World


A prototype in a sandbox is like an orchid in a greenhouse: It thrives only as long as the conditions are right. A sandbox exists to test an idea, not to build a working business application. An artificial intelligence (AI) model may work perfectly when the problem is narrowly defined and the data carefully curated. Functioning in a chaotic, unpredictable world is another story.

Think about your own experience. When a robotic voice last asked you to “explain in a few words” what you needed, were you understood? Did you have to speak with a human? Did it even send you to the right human?

“People don't think about how difficult it is to follow a conversation,” says Lili Cheng, corporate vice president of Microsoft’s AI and research division. “We’re still learning many of the best practices for designing a clear and successful experience.”
No application of any type, AI included, can deliver on its business promise until it moves from a lab prototype to a robust, real-world tool. Many AI models do pass that test. But often only after a long, difficult journey—one with pitfalls that could have been avoided.
uncaptioned
Lili Cheng, corporate vice president, Microsoft AI and research divisionMicrosoft

How Models Move Into Production

The development process often starts slowly and speeds up over time. Data scientists and developers come up with a prototype, test it, modify it based on what they learn and test it again. The feedback-test-redesign process repeats itself until the prototype works well enough to be used more widely. 

A big part of the testing is seeing how the model responds to real-world data. “When you start, you might not have much data, but you get a lot more once you start testing the application and see what people actually do,” says Cheng. “It’s important to get the solution prototyped fast so you can get feedback in the form of user data and understand where the problems might lie.”

Start With A Strong Problem Definition

Most important is the problem definition—and that’s a process that needs to involve everyone. “You start by talking to a broad group of business stakeholders to make sure you're solving the right problem with AI,” says Cheng. Then comes setting expectations for the solution—what’s possible and over what time frame—and a plan of action.

A big part of setting expectations is educating the business on what AI is and isn’t. A user may see what looks like a widget on a webpage. But on the other side of that door is a complex system that interprets data and draws conclusions based on algorithms that mimic what a human would do. If the application looks simple, it’s probably because it is designed well, not because it actually is simple.

There Are No Design Conventions

As the internet evolves, design conventions take shape and users come to know what to expect. The design principles of webpages, for instance, are well-established: menus on top, a search bar, perhaps a navigation sidebar.
But there are no user-experience standards for bots. “Users try all kinds of things and get frustrated,” says Cheng. “You have to make it clear what the app does, but you don’t have a standard way to show them. We're still learning best practices for how to design an experience that is clear and successful.” That’s why teams need to pay attention to the basics of the user interface before turning to advanced functionality.

Pre-Built vs. Custom AI Models

The speed of development often hinges on whether the organization starts with a pre-built AI model or attempts to create a completely customized model from scratch. Most organizations start with pre-built models because they speed up the development process. Speed is important because “it's critical that companies are able to train their models fast, have users try them out in their use cases and then iterate quickly to test the AI application,” says Cheng. “Then you just need a continual feedback loop.”

Moreover, starting with pre-built doesn’t prevent customization down the road. Most natural language processing applications have custom dictionaries, for instance, as well as exceptions built around specific technical or customer terms. And that kind of customization isn’t hard to do. “We can make it easy to add words or otherwise customize a model without having to get into all the details,” says Cheng. “Someone might see errors in a log and say, ‘Oh, let's add this term to the dictionary.’ There’s no need to know how to code to retain the model or even to have much technical knowledge.”

Pre-built models are multiplying and improving: “If you're trying to do something very specific and have a good data set, you’ll do something that's more custom,” says Cheng. “But as people create more models with bigger data sets updated more often, the pre-built tools get better and better. My rule of thumb would be: Before you make a big investment in a custom model, especially if you lack data, see how far you can get with the pre-built models.”

Choosing How To Work With Outside Experts

A prototype might be built by a few technical specialists, but bringing that prototype into the world takes a village. It requires contributions from data scientists, software engineers and designers, as well as oversight by the business—a mix of skill sets and perspectives from inside and outside the business.
An operational AI application can be developed in three ways:
1. Developed, owned and maintained by a single organization
2. Built by an outside specialist organization and deployed and run internally
3. AI-as-a-service: Built, hosted and run by an outside organization
Below, we look at each of these in more detail.

Build and deploy
A company with a strong core of developers, where the business wants to own the AI from the conceptual level all the way down to the code, might do it all in-house. “We might get involved by organizing an internal hackathon,” says Cheng. “Once the developers have worked with us for a week, they have a solution that they can continue to build in their own company.”

Deploy a system built elsewhere
Companies without the AI or development skills might ask an outside organization to build an AI application that it then deploys and runs. These companies tend to have a pipeline of good data that they want to use over time, as well as the ability to maintain the system. “Our self-serve tools are available on GitHub, which any developer can use,” says Cheng. “There is no need to interact with us unless they choose to.”

AI-as-a-service
Here an outside organization builds, hosts, runs and delivers the AI application. The most common category of AI-as-a-service is natural language processing, followed by vision services such as facial recognition. 
Choosing the right development approach is essential for success. The choice depends not only on the in-house skill set, but also on the scope of the project and the urgency of the timeline. Most companies don’t have sufficiently strong in-house AI skills, so the latter two models are more common.

Who Do You Want To Be?

To a generation that has grown up with the internet, the revolutionary promise of its early days—or even the early days of the internet’s second wave, which brought social, mobile and cloud—may seem tired and outdated. But AI truly is something different—something that holds tremendous promise for changing the way organizations run.

“The concept of a web browser or the mobile phone didn’t trigger conversations like the ones we’re having now,” says Cheng. “People are asking questions like, ‘What is the future of the company?’ and ‘How do we want our customers to engage with us?’”
“It’s so cool for us to be able to have these kinds of conversations,” she says. “Just the phrase ‘artificial intelligence’ has the effect of setting ambitions to a higher level. It raises the question, ‘Who do you want to be?’ And that ambition means that bigger and better things will happen.”
Learn more about how companies are leveraging AI today.
CREDITS: Akrain/iStock



About the Author
Forbes Insights is the strategic research and thought leadership practice of Forbes Media. By leveraging proprietary databases of senior-level executives in the Forbes community, Forbes Insights conducts research on a wide range of topics to position brands as thought leaders and drive stakeholder engagement.

No comments:

Post a Comment