MIT SMR Connections
MIT SMR Connections is an independent content creation unit within MIT Sloan Management Review. We develop high-quality content commissioned and funded by sponsors. We welcome sponsor input during the development process but retain control over the final product. MIT SMR Connections operates independently of the MIT Sloan Management Review editorial group.
In this Q&A with MIT SMR Connections, Michelle K. Lee, vice president of the Amazon Web Services (AWS) Machine Learning Solutions Lab, shares real-world examples of machine learning in action, describes four key implementation challenges, and offers other advice.
This conversation has been condensed and edited for clarity, length, and editorial style.
MIT SMR Connections: Can you provide an overview of how artificial intelligence (AI) and machine learning (ML) are driving digital transformation?
Lee: AI and machine learning went from being aspirational technology to mainstream extremely fast. For a long time, that technology was limited to a few major tech companies and hardcore academic researchers. But this began to change with three primary advances in technology.
First is the increase in the power of computers. Second: the decrease in storage price. And third is cloud computing. Machine learning requires extremely powerful computers to pore over large amounts of data that is easily accessible. Through cloud services pioneered by AWS, the powerful computers and access to the large amounts of data necessary for ML are now readily available to all, not just a few major tech companies and academic researchers.
As a result, almost every industry — finance, retail, agriculture, health care, manufacturing, and, really, every business — has the opportunity to take advantage of the recent advances in machine learning. When I talk to executives now, they are no longer asking, “Why should I be looking to employ machine learning in my business?” but instead, “How should I go about doing so, and how can I be successful in that?”
MIT SMR Connections: Could you provide examples of businesses using ML for forecasting, prediction, and decision-making?
Lee: Domino’s Pizza is using Amazon Personalize to predict purchasing behavior and then delivering personalized promotions and notifications to their customers via digital channels, including their popular mobile application. So instead of sending the same SMS [text] promotions to everyone, they can send them at times that suit the customer with the content that is most likely to result in a conversion of a purchase.
Intuit has employed ML-driven forecasting capabilities to predict call-center demand on a particular day or time to make sure that their customer service agents are adequately staffed.
In the health care space, we are helping companies to make better, faster decisions. We’re seeing a shift from reactive to predictive care, including the use of predictive models to accelerate research and discovery of new drugs and treatment. Cerner, one of the largest publicly traded health care IT companies, used Amazon SageMaker to build a solution that enables researchers to query anonymized patient data records to predict congestive heart failure up to 15 months before clinical manifestation. They’re also using Amazon’s Transcribe Medical to free physicians from the tedious task of note-taking by providing a virtual [voice-to-text] scribe so the doctor can focus more on the patient interaction.
Convoy, a Seattle-based logistics company, is disrupting the trucking industry by using machine learning to more efficiently match truck drivers wanting to drive loads with shippers needing to transport loads, resulting in lower costs and faster deliveries. As a result, truck drivers are driving more often, according to the schedules they want, and those who need their goods transported are able to connect with drivers more easily through the machine learning solution.
Another example, from a completely different industry, involves the National Football League. The NFL developed a set of next-generation stats and ML models using its historical data on, for example, pass completions. Then, when the fan is watching the game on TV, the NFL provides stats about the probability that the player will actually catch the football for a completion as the play is unfolding. These almost instantaneous predictions enhance the fan viewing experience and are enabled by the ML models combined with real-time data from the field.
MIT SMR Connections: Those are great examples. But how do companies with less experience know whether they need ML in the first place?
Lee: I believe that every organization has a machine learning opportunity.
What business wouldn’t benefit from improved data-driven forecasting of customer demand, or enhanced forecasting of supply chain and inventory needs? What business wouldn’t benefit from enhanced personalization of services and products offered to customers to help drive revenue? What business wouldn’t benefit from automation in the very labor-intensive customer-call centers? Or from a quick assessment of the customer sentiment about the company, its performance, or its product offerings through online reviews, social media, or even by recordings from customer call centers?
I had the privilege of leading the U.S. Patent and Trademark Office, a governmental agency that has been examining patent applications pretty much the same way for over 200 years. Because I have this artificial intelligence and computer science background, I recognized that I could use data and data analytics to improve the quality and consistency of the patents issued, so I implemented that at the USPTO. I’d say if a 200-year-old governmental agency has a machine learning opportunity, I would imagine most businesses probably do, too. It’s just a matter of finding it.
MIT SMR Connections: Are there misconceptions about ML from a business standpoint?
Lee: Probably the biggest one is that the only thing standing between you and your dream ML application is a team of data scientists. In actuality, a number of factors need to come together in order to achieve a successful machine learning implementation.
Yes, you do need data scientists, either on your team or as consultants. But, equally important, you need to identify and tackle the right machine learning use case for your company — one that solves a real and significant problem that has measurable return on investment. You also need to have the data required to support the building, training, and testing of your machine learning model. And it certainly helps to have senior-level business buy-in for the project so that it is not simply a science experiment but something that solves a real business problem and is incorporated into the fiber of the business.
MIT SMR Connections: What kinds of ML specialists do companies need today?
Lee: Particular roles that are necessary include data scientists, data engineers, software developers, and technical program managers. A variety of skills are needed, and the key for a company is to do a skills analysis to identify the gaps up front. Data analytics and machine learning, at least in their current forms, are relatively new disciplines, so there is a shortage of people with these skills. This means that a company probably isn’t going to be able to hire all of them, so perhaps it ought to focus on training its current workforce.
At Amazon, we took an approach to both hire new talent and develop existing talent. Amazon developed a machine learning university that we have used for over six years to train our engineers. Last year, we made a lot of this content available for free to our customers — and, actually, to the public too. We have seen well over 100,000 developers start their machine learning journeys using this content.
MIT SMR Connections: What are some common challenges that companies may face in adopting ML?
Lee: We’ve learned four key challenges that leaders need to address for successful adoption of machine learning: data strategy, getting started, the ML skills gap, and spending time on undifferentiated heavy lifting. That last challenge refers to activities such as building their own infrastructure and tools for data aggregation, access, and cleanup and modeling, rather than taking advantage of existing services such as Amazon’s data lake offering, SageMaker for helping with ML model building and deployment, Rekognition for computer vision, Translate for language translation, or Comprehend for natural language processing.
Data is often cited as the No. 1 challenge in adopting machine learning. To be successful in machine learning, a company needs to have a data strategy that identifies the data it has, where it’s located, who controls it, and where it needs to be to support its full and optimal use by the company. A company also needs to ask, “What data don’t I have today that I want to have in the future?” and then begin developing a plan to gather such data.
Without a data strategy, the ML scientists a company hires will spend an inordinate amount of time dealing with data-management access and cleanup or, worse, get bogged down and frustrated because they lack what they need to solve the larger problem. So companies need to enable the IT team to break down any data silos and to collect the right data in a safe and compliant way.
A second challenge is, how do I get started? Although every business has a machine learning opportunity, not every business problem is solvable by machine learning. So identifying that high-value use case whose results are measurable is key. But that’s not always easy. There is a lot of hype around what machine learning can do. That’s why AWS created the Machine Learning Solutions Lab, which allows us to work side by side with our customers, to listen to their business problems, to identify their highest-value ML use cases, and to help guide them to implementation. To each of our engagements, we bring tremendous depth and breadth of experience and expertise based on our engagements across a wide range of industries and use cases.
The third challenge is the skills gap. Again, the growth in artificial intelligence has led to a shortage of data scientists and machine learning experts. You may not be able to hire all the data scientists you need, so you should probably focus your energy on upskilling the level of your current workforce and/or leveraging outside resources.
And a fourth challenge is the tendency to think you have to develop everything on your own from scratch, when a cloud platform like AWS can provide many of the necessary tools and infrastructure needed for data access and machine learning model development, testing, and deployment. By taking advantage of these existing tools and services, you can focus on bringing your differentiated, value-added contributions, such as your domain and industry expertise and any special insights that you have, to solve the problem at hand.
MIT SMR Connections: What kind of culture do organizations need to succeed with ML?
Lee: Machine learning requires a cultural shift that’s most successfully driven from the top. As a leader, it’s important to articulate the priority of machine learning to the company and to encourage team members to continually ask themselves whether a business problem might be better solved with machine learning. Again, not every business problem is best solved by machine learning. But constantly asking that question is critical.
Ten years ago, the Amazon leadership team asked every business leader at Amazon — regardless of whether they were running a research team, a fulfillment center, an HR organization, or the legal department — how they planned to leverage machine learning in their business. “We don’t plan to” was not an acceptable answer. This forced every part of the organization to think about how ML could improve some aspect of their business and to develop a plan to achieve it. Today, I would say there’s not a single business function at Amazon that isn’t made better through machine learning.
But this didn’t happen overnight. It took a cultural and a technological shift.
MIT SMR Connections: What else would you like business leaders to know about ML?
Lee: Machine learning is still in its infancy, but it’s not entirely new. Still, the path to machine learning success is not always straightforward, so many organizations need a partner to help them along the journey.
We have successfully helped so many organizations, from Domino’s Pizza, to the NFL, Cerner, and NASA, achieve machine learning successes.
While we always aim to help our customers identify and deploy their high-value ML use cases, our goal is also to teach our customers “how to fish.” To this end, we offer a program called AWS Machine Learning Embark, which not only provides workshops and ideation sessions to help identify their best use cases, but also machine learning training for both technical and business leaders for the precise reason I mentioned earlier: You want people within your organization, at every level, to be thinking, “How might machine learning improve or solve the business problem at hand?”
As vice president of the Machine Learning Solutions Lab at AWS, Michelle K. Lee leads a global business focused on helping AWS customers identify high-value machine learning use cases and guiding them to implementation. Previously, she was the under secretary of commerce for intellectual property and director of the United States Patent and Trademark Office, an executive at Google, a partner at the law firm Fenwick & West, and a computer scientist at Hewlett-Packard Research Laboratories and the MIT Artificial Intelligence Laboratory. She received bachelor’s and master’s degrees in electrical engineering from MIT and a J.D. from Stanford Law School.