Capital One Managing Product VP on Building Smarter AI Products (PART 3)

To understand AI/ML, it helps to know about unsupervised machine learning. When looking at AI products, unsupervised learning works when grouping similarities. For example, Google Translate leverages unsupervised learning thanks to the huge amount of data it taps into.

Unsupervised machine learning can take a bunch of images and group them, let’s say into dogs and cats. However, it can’t figure out the name of the group on its own. It would be able to assign new images, if they match, into established categories.

Unlearn What You Have Learned

Any AI/Ml model must get integrated into your general software system. What if you want to add features later? You’ll need to retrain the model. Returning to the animal example, if you want to introduce a mouse category, you have to be careful not to undo prior learning. Still, in some ways, intentional unlearning is important.

As the machine learning progressed for our intelligent assistant Eno, we encountered the challenge of intent splitting. When someone asked, “What’s the status of my credit line increase?”, we answered, “To find out this information, please go here.”

But what if they wanted to apply for a credit line increase? Our AI/ML model wasn’t ready for the subtle word difference but big difference in meaning. So we had to retrain the model using data examples now split into two categories.

Industry Lessons

Decide quickly if your internal team can handle the job, or if you need to bring in a third party. Currently, ML experience is scarce, and you need deep pockets to get good talent. Think carefully about how much you can invest, especially as you move past experimental phases.

There’s also a lot of traditional software engineering that must accompany new AI/ML models. For instance, a bank chatbot needs to make API calls to get customer data. Finally, deep data understanding is critical. How much data do you have? What are its limitations? What’s missing, and how does this affect edge cases?

The Future of The Industry

In traditional digital development, we’ve been trained to point, click, fill out forms and hit submit. In a conversational model, I don’t get to decide how my customer asks a question. True AI must be able to interpret and go with variations. It’s an entirely new paradigm, and this is a huge area of future AI/ML research, development and product design.

For instance, Eno can understand 2000 different ways customers ask for their balance which includes misspellings, variations and abbreviations. Humans generally get all that naturally. Machines have to learn it.

Click here for Part 1

Click here for Part 2

Comments

Please to leave your comment

Load more comments
About the Speaker
Margaret Mayer
Capital One Managing VP
Margaret Mayer is the Managing VP of Messaging and Conversational AI at Capital One - setting the technology strategy for the company's banking chatbot (Eno) and enterprise-level messaging platforms. Over the past two decades, Margaret has held several leadership positions at Capital One - including software engineering, consumer identity and partnership development. Margaret holds a PhD in industrial engineering from Lehigh University and a bachelors degree from Cornell University. She currently lives in Richmond, VA.

Recent Posts