Deploying AI instantly brought value and growth to many businesses. However, it is well established that sustaining the value over time, not to mention maximizing it, could be quite challenging. Continuous optimization is the key to successful AI deployments. Beginning with a product that’s good enough, learning from how it performs in the real world, especially as the world (read: the data environment) changes, and then improving; then learning and improving again and so on. It’s a bit of an obvious insight but it is rare for AI-driven products to be perfect from day one.
To accomplish continuous optimization you need continuous feedback. You need “eyes and ears” observing your data and models and telling you whether they’re performing well. That’s easier said than done, for various reasons. We have outlined these reasons below.
Before we go there, let’s first establish what we mean by continuous feedback. It is bridging the gap between the business outcomes and AI workflows. Are you able to demonstrate how your AI is impacting business KPIs? How about how changes in your AI are impacting changes in the business KPIs? If there are changes in business KPIs that are rooted in changes in your AI, can you detect that? If the answers are YES, then you have continuous feedback.
Now, back to the challenges of attaining continuous feedback for your AI.
1. Defining the desired outcomes
First, you need a clear definition of what success looks like. And it’s not always obvious. For example, two recommendation systems may look very similar but in fact, they have very different strategic objectives. On one hand, the goal is to drive immediate conversions, like you would on an eCommerce website. On the other hand, it is to drive general customer satisfaction, such as content suggestions from a streaming service. It’s not always the easiest, but it is critical to measure your AI in general, based on indicators of the main objective.
2. Disparate systems
So, once the business indicators (of the desired outcomes) are well defined, you need to track them and repeatedly analyze them in the context of changes in your AI (data, features, inferences etc). A common challenge around doing that is that the business indicators “live” in a different environment, outside of the AI stack. For instance, in an eCommerce recommendation system, the business indicators may be clicks and conversions and these are tracked in your marketing stack. Can the AI team access these indicators easily and consistently for continuous feedback? In our experience, not very often.
In some cases, the business indicators are well defined, and the AI team can access them, but the challenge is with timing. This means that the indicators are measured at a later time, and are not available when (or shortly after) your AI is at play. An example for this is credit models. A prediction of an individual’s credit worthiness is used to automatically approve a loan or line of credit. Then, it could be months or even years before the accuracy of that credit risk prediction can be determined. In other cases, the lag time may not be as extreme (maybe just days or weeks), so it will be worth designing a mechanism and implementing an integration that will provide that feedback once the data is available. It is important to plan for timing challenges in which it can be mitigated through partial feedback (a subset of data to feed the system), manual feedback (adding a human into the loop), or proxy feedback (e.g., shrinking confidence scores).
The last challenge that we will outline for continuous feedback is that in some cases it can be very expensive to obtain the indicators of success. What we mean here is that often the indicator of success requires a human in the loop. An annotator to provide a “ground truth” label. Labeling data may require expertise (e.g., medical expertise to verify the read of an X-ray) or simply work hours (going through a million images), either of which could prove costly.
Continuous feedback will improve your AI over time and accelerate business value, give you confidence in your data pipelines, inference engines, and your model environment overall. While it’s not always easy to achieve, forward thinking AI teams see continuous feedback as a critical element of production grade AI and work to institute it, often through an intelligent monitoring solution. Mona is the leading platform to bridge that gap between your AI workflows and business outcomes. Discuss how you can properly connect your AI data with feedback data by contacting us or you can try out our monitoring platform by using our tutorial with a free account.