If You Can’t Trust a Machine, Then Who Can You Trust?
As machine learning becomes pervasive and our usage of this technology expands here at SAP, we’re thinking more and more about how it relates to trust by our end users. When a system gives you an answer to your business question,what allows you to trust this answer and therefore get value from it? Our belief is that reliability, transparency and consistency are required to trust a machine learning powered system, and so we choose to build these qualities into our products alongside qualities core to machine learning like robustness and accuracy.
Humans Have Established Ways of Building Trust
Why these qualities? Because these are the same qualities we desire in people. We naturally build trust in someone over time when they do what they say they will do, they do it well, and they do it consistently.
When we meet someone for the first time as a complete stranger, we have no basis to know how they will behave. So for example, if we are on the same football team and they offer to give us a ride, we will probably instinctively create a backup plan as we don’t know if they will show up. After that first ride, we become less cautious as we know they intend to follow through. Of course, if that same friend drives erratically or gets us lost, we will lose this confidence in them as clearly they lack skills. Lastly, the friend has to continue to do what they have promised, otherwise we will lose trust.
Trusting Augmented Systems
Why though, is trust important for SAP systems when it comes to machine learning? To build trusted features, we have to embed the qualities that people use to build trust with each other into our systems. As SAP, this means we are investing in a multitude of aspects to build the trust between our end users and our systems.
Priority 1: Be Clear
First, it means transparency, and clarity of why we are giving you a particular answer. We do this through various techniques of Explainable AI including reason codes, key influencers, quality and precision KPIs when using machine learning in the background.
So for example, in SAP Analytics Cloud, if you predict customer lifetime value, the system automatically provides a list of key indicators that contribute to the prediction as well as a list of customers that were so unusual that they were excluded from the underlying machine learning model.
Priority 2: Be Consistent
Second, it means being consistent. At SAP, this means being able to provide a similar experience across a wide range of use cases.
For example, being able to forecast in different workflows throughout SAP Analytics Cloud and seeing the same quality indicators whether you are working inside a chart, building a planning forecast, or building a predictive model. With the same data, you should receive the same answer, regardless of the workflow.
Priority 3: Be Reliable
Third, it means being reliable. The most important part of being trustworthy is to tell it like it is, one of our core SAP values. Sometimes a machine learning algorithm can’t solve your problem, as there isn’t enough input data, there aren’t any strong patterns in the data, or the data is not in the right shape. These scenarios are highlighted by low quality indicators, or if it isn’t reliable we will give you no results at all.
This gives you the confidence in our respect for model accuracy and robustness that when you do get results you can trust them.
IDC trends show that data will explode ten-fold by 2025, meaning human intelligence simply isn’t enough. Without being able to trust automated systems, you can’t harness the value of the vast data inside.
What do you need in order to trust a system to provide you an automated answer?