Activating ML in the Enterprise: An Interview with Michelle Lee, VP of Amazon Machine Learning Solutions Labs

In the previous blog post I explored with Michelle K. Lee some of the societal impacts of artificial intelligence (AI) and machine learning (ML). In this post I dive into the patterns Michelle has seen organisations implement to take advantage of the promises of ML.

―Phil

Some surveys show a gap between an understanding of the value of AI/ML as a transformative tool and a practical understanding of what that means and how it can be prioritised. How effectively have enterprises taken advantage of the promise of these tools?

In reality, most of us are using AI- and ML-based systems multiple times every day. For instance, when you deposit a check by taking a photo of it, a machine learning model developed by your bank recognizes what’s written on the check. This prevalence of ML in our everyday lives is evidence that enterprises are taking advantage of the promise of this innovative technology—but it doesn’t mean that there isn’t more work to be done to bridge the gap between understanding its value and making it a business priority and asset. The Amazon Machine Learning Solutions Lab has helped countless customers adopt ML, and the most successful engagements have the following characteristics: an ambitious, top-down vision for applying ML across an entire organization, and a mandate that ML be considered a business priority. Without these, it’s difficult to move beyond understanding that ML is powerful and transformative to benefitting from it.

Where are the more significant gaps in application? What are the most common barriers?

As recently as five years ago, the most common barriers were technical, and by extension, cost related. Only large, technically sophisticated businesses could afford to acquire the technology infrastructure and human capital to apply ML in even a modest way. Cloud computing has eliminated those barriers. Now, companies of any size and level of expertise can use ML in an expansive way. The most common barriers we see today are business and culture related. Defining and executing an ambitious vision for ML is hard and takes time. Developing machine learning applications is an iterative process, requiring experimentation. If your organizational culture doesn’t encourage experimentation or if it treats failure (AKA learning) as something to be avoided at all costs, then this will be a significant barrier to applying ML effectively. What’s interesting and encouraging is that even in cases where the cost of getting something wrong is unacceptably high, we are still seeing widespread adoption of machine learning. In these cases, the need for experimentation is even higher. The more experiments you run, the more you learn about the problem space you’re exploring. So while some types of businesses or industries may be adopting ML more slowly than others, we are seeing quite a bit of breadth in the use of this technology.

How can companies best address questions around talent and infrastructure and the use of AI/ML?

While cloud computing has solved much of the infrastructure problem, you do need to align your technical talent with the right set of tools. This is why we organize our ML services into a three-layered stack. As you move from top to bottom of the stack, the level of required ML expertise increases. As companies begin their ML journey, they need to have a plan for reskilling their technical talent and aligning this talent to the right set of services. Companies which are new to ML, but have software engineers with experience in AWS, can get started immediately with the top layer of services. As engineers develop deeper ML expertise, then it’s appropriate to explore the next layer of the stack. At Amazon, we’ve enjoyed quite a bit of success in training our engineers in ML, and we recommend this approach. You can certainly build a team of ML experts from scratch, but we think there’s value in investing in your current technical talent and providing avenues to expanded job opportunities.

Many companies feel the need to be seen to be investing in AI, but often don’t know what they don’t know. How would you recommend getting senior, non-IT leadership better informed about AI and how it fits with their business objectives?

At Amazon, we have over twenty years of experience using AI and ML across a wide range of our businesses. And we’ve learned a lot along the way. We’ve had to train both our business and technical leaders in ML, and we’ve developed a curriculum for delivering this training internally at Amazon. Two years ago, we made the technical portion of this training available as part of Amazon’s Machine Learning University program. A little over a year ago, at our re:Invent conference, we announced the Embark program, which combines the technical training from Machine Learning University with a new curriculum designed for business leaders.

One of the primary goals of the Embark program is to demystify machine learning. Upon completion of the Embark business training, senior leadership will know the types of problems that ML is well suited for, and how to identify areas of their business to which ML can be applied. To do this, we discuss real world examples of where we’ve seen machine learning work well—both for Amazon and the more than 100,000 customers using ML on AWS today. They also get exposed to various challenges in developing an AI-powered organization. Through this, they learn to identify the strategic levers to creating a robust AI strategy, data strategy, and culture of collaboration, required to successfully shift their organization toward AI and ML. We also make the benefits of ML tangible to the customers participating in the Embark program. Data scientists from the Amazon Machine Learning Solutions Lab will work with the customer’s team to develop a fully functional ML proof-of-concept addressing an actual business use-case we help the customer identify and define. Then the customer can implement to production using AWS’s ProServe or customer’s developer team, at their option.

From your experience, what do you see as the critical success factors to consider before starting a journey with ML?

One of the most important points for business leaders to consider is that the challenges they are likely to face are nontechnical; that is, these challenges are mostly about organization and people.

The first of these is setting a data strategy. Collaboration between your technical and business experts is going to be far more productive if your framework for decision making is data driven. This is absolutely critical for ML, because at its core, ML is all about finding patterns in data. The details for designing an effective data strategy are far too numerous to cover here, but are covered in some depth as part of the Embark training. Considerations include ensuring access to data is controlled (but not prohibitively so), ensuring that both raw and derivative data is available, and expecting to leverage multi-modal data (e.g., structured, semi-structured, and unstructured) with ML.

For example, let’s say you want to put together a special offer, sent over email, to your customers. The traditional approach would be to rely on historical data about which offers customers have responded to in the past. This data probably sits in a database somewhere. With ML, you can leverage different types of data like images, and then use computer vision to find images that are similar to ones that have been used in offers your customers have engaged with in the past. This is what I mean by “semi-structured” and “unstructured” data. There’s a whole universe of data out there that doesn’t fit into traditional databases, and ML is extremely good at finding insights from this data.

The second challenge pertains to talent. You’ll want to generate excitement about ML so that your engineers are enthusiastic about building their skill sets. You can certainly hire data scientists and ML engineers, but these skills are in high demand and short supply. We’ve found that it’s both faster and more effective to up-skill your existing technical teams. This doesn’t mean you won’t have to hire engineers with new skill sets; it just means that you won’t have to hire an entire team from scratch. Amazon’s Machine Learning University is based on the training curriculum we developed internally to up-skill our own workforce, and which is now available as part of the Embark program as well. We’ve enjoyed quite a bit of success in training our own workforce, and we believe that other companies will too.

The next challenge is that you must identify impactful business problems that are good candidates for ML. This is where your business domain experts can provide valuable insights. Early on in your ML journey, you want to identify projects with a high probability of success, but that have enough business value to be meaningful. These projects should deliver results that cause leaders across the organization to say “Wow, that’s impressive. I’ve got some business challenges in my department that I’d like to solve with ML too.” We cover how to do this extensively in the Embark program business training as well as our ML Solutions Lab engagements with customers, where we intentionally bring together the customer’s business leaders and technology leaders to jointly identify their highest ROI business problems to tackle through ML. This collaborative approach ensures key stakeholder buy-in and alignment with business goals from the very start, crucial to ensuring a successful outcome and adoption.

The final key challenge is scaling out ML within your organization. The technical side of running ML at scale is solvable through the AWS cloud’s services and solutions for scaling out technical infrastructure in a highly cost-effective way. There’s a business component to this challenge as well. Once business units see how powerful and transformative ML is, they’ll want to implement solutions in their lines of business. You’ll need to give some thought how to scale demand for your ML engineering resources. As you get started, consider a centralized ML and data science team that cuts horizontally across the organization for data collection and a decentralized team embedded in the business for identification of use cases as well as building, testing, and deploying the models.

There’s no single “right” approach. It’s tempting to be completely prescriptive when addressing challenges like these, but every customer is different. So what we do with our Embark program is make business leaders aware of the key decisions they’ll need to make as they begin their ML journey. We’re committed to making ML easier to adopt and use, and we think our Embark program represents a big step in advancing that goal. If you want to learn more about our Embark program, please reach out. We’re excited to work with you!

About our Guest

Michelle Lee, Vice President of the Amazon Machine Learning Solutions Lab at Amazon Web Services

Prior to joining Amazon in 2019, Lee was an MIT CSAIL (formerly MIT Artificial Intelligence Lab) computer scientist, a tech executive who helped build a company that, like Amazon, grew quickly into a multi-national corporation, a professor, and a public servant. In the latter role, she served as the Under Secretary of Commerce for Intellectual Property and director of the United States Patent and Trademark Office (USPTO) from 2014 to 2017. In this position, Lee was the chief executive officer of one of the largest intellectual property offices in the world with around 13,000 employees and an annual budget in excess of $3 billion, and served as the principal advisor to the president on intellectual property policies.  She is the first woman to hold this position in American history.