What is your AI really doing? VR could unlock its secrets.
3D visualizations offer a fresh perspective.
You’ve finally found your dream home and apply for a loan at your bank. You’re busy picking out paint colors and thinking about a kitchen remodel when you get a notification: your loan application has been rejected. Wait… what? You have a pretty decent credit score, a good job, so this makes no sense. What’s more, the bank can’t seem to explain the decision. It turns out a machine learning model was involved in the process and things are complicated…
If you asked 100 people to define AI, chances are you’d get 100 very different answers. But if there’s one thing we can all agree on, it’s that the more complex AI becomes, the more enigmatic it gets. It’s what we call the “black box” problem — we put some data into our AI model, some magic happens in the black box, and a result comes out the other end. These results help us automate, diagnose, or predict so many different things that truly impact our lives. From healthcare, to finance, to transportation — AI is all around us. Yet we still don’t really understand it. How do we demystify AI? How do we start cracking open that black box and begin making sense of what’s inside?
We know that data visualization is key for exploration, but 2D visualizations don’t quite cut it when it comes to large, multi-dimensional data sets and machine learning models. At IBM, my colleagues and I are working on a new way of exploring AI models using augmented and virtual realities. By stepping away from 2D screens and into a 3D world, we’re giving users a better understanding of how and why an AI model works. Users can interact with their model spatially, see how features in their machine learning model relate to one another, understand which variables in the model are most important — and ultimately, shed light on AI’s decision-making process.
How do 3D visualizations help?
Machine learning models range in complexity. On the one hand, you have simple models that are generally easy to comprehend but tend to be less accurate. On the other hand, you have complex models that are much more precise, but we don’t understand the logic behind their predictions. This means data scientists constantly have to make a trade-off between interpretability and accuracy. By giving data scientists more intuitive tools to make sense of AI models, we hope they’ll no longer have to make this compromise.
Using IBM Immersive Data, you can step into your AI model in virtual reality. You can start to get a sense of which features within the model are most important, as well as what dependencies a particular feature has. You can also select some features to reveal a 3D scatter plot which shows two variables and the SHAP value. By exploring your model this way, you can start to get a better sense of the contribution of a feature to the target, and the interaction between two features. The extra dimension in AR/VR allows us to create new types of visualizations that take our understanding of the model to the next level.
What would this look like in action?
Imagine a bank wants to quickly explore a credit loss model as required by CECL (Current Expected Credit Loss) regulations. A lot of different factors can impact the quality of a loan. These could be specific to the loan (for example, contractual terms, the loan amount, or amortization schedule) or they could be external (e.g. unemployment rates or interest rates). The bank needs to collect a lot of data — including many years of historical data — to create a long-term view of risk. However, when you have so much data, finding a starting point for exploration can be overwhelming. IBM Immersive Data works in conjunction with other tools to bubble up interesting insights and speed up the process. It then allows you to step into the data in VR and view it in a whole new way. Check out the video walk-through showing this exploration using a real public domain loan data set.
Coming up with an accurate loss prediction model is important to the bank because it increases revenue, reduces the risk of losses, improves the customer experience, and importantly, ensures compliance with regulation. Lenders need to be able to explain to customers why they were rejected for a loan. Despite the complexity of the model, IBM Immersive Data allows the bank to tease it apart and begin understanding it.
As you can see, this is still a very new technology that is evolving, but the potential is tremendous. By being able to see inside the black box, users are able to quickly explore and understand how their machine learning models are making predictions. This allows them to then take other steps such as choosing a different model for their task, fine-tuning their model, mitigating bias, or providing greater transparency into their processes.
In addition to interpreting AI models, the Immersive Data visualization tool can be used to explore highly multi-dimensional data sets. This allows users to quickly spot patterns, correlations, and outliers by seeing more dimensions of their data at once. The volume of data businesses are grappling with is increasing exponentially as IoT, edge, and other “smart” devices become more and more commonplace, so the ability to distill this data into meaningful insight will drive growth and value. We aim to offer a way for organizations to not only visualize their data and algorithms, but also to share those results and collaborate with their teams and clients in a more intuitive and engaging way. Ultimately, our goal is to develop IBM Immersive Data into the visual analytics tool of the future.
Reena Ganga is Design Lead, AR/VR at IBM’s Silicon Valley Lab. The above article is personal and does not necessarily represent IBM’s positions, strategies or opinions.