The core focus of the Optimizing Data-to-Learning-to-Action method is learning. We use that word all the time, and we have a common intuition of what it means when we say that we are learning a new skill or that we have learned about what a friend was doing on Tuesday night. But when we want a concept such as learning to be extensible so that it can effectively combine with other concepts to form a new methodological whole, the vagueness of general intuition is insufficient–we need to unpack the concept and determine its essential characteristics.
These simple examples above are, of course, examples of human learning, which from an organizational learning perspective has historically been the only type of learning that needed to be addressed. However, especially in view of the rapidly growing importance of machine learning in organizations, we really need to generalize by coming at the question of what exactly learning is most fundamentally from two different perspectives—learning as it is performed by people and learning as it is performed by machine. Fortunately for our method the answer converges on a common understanding of the essential characteristic of learning.
What we necessarily conclude from examining both perspectives is that learning most fundamentally can be considered the process of better predicting, and that better predicting is achieved through the reduction of uncertainty about what is being predicted. This is certainly the way machine learning works, for example. A learning system makes predictions (such as what an image represents, or what will happen when a robotic leg is extended) and then it (maybe with our help) assesses how these predictions compare to what actually does happen after it performs an inference or action. Future predictions are then adjusted based on this input, with the objective of continuously shrinking the predictive errors.
It is increasingly understood that this process is also most fundamentally the way that human learning occurs. The same type of prediction-to-assessment-to-learning process (where the prediction might simply be called an “expectation”) occurs as in machine learning. And frankly, that certainly should not be surprising–why shouldn’t we expect that, say, an artificial neural network, and our flesh and blood neural networks at the most basic level necessarily learn in the same way?
Whether with respect to minds or machines, the uncertainty that results from the gap between predictions and reality can be encoded in probabilistic form. Such a probabilistic-based form doesn’t necessarily have to be explicitly applied, although it sometimes indeed is for both human-based and machine-based learning. But regardless, implicitly, if not explicitly, the predictive uncertainty can be fundamentally considered probabilistic in nature and learning therefore is essentially the process that causes changes to probability distributions. And in some cases, there will be a chain of changes to probability distributions (i.e., learning) that flows forward to affect other probability distributions.
Although it can seem a little abstract to apply this definition to our everyday examples above, it is readily seen that the underlying concept indeed applies. When we learn a new skill, we have less uncertainty about that skill area. When we learn something about what our friend did on Tuesday night, we have less uncertainty about exactly what our friend was doing during that time.
Now, again, this perspective on the essential nature of learning applies in a consistent fashion to both minds and machines. That has big advantages as we apply the optimizing data-to-learning-to-action method in this era of rapidly advancing machine-based learning because it provides us with an apples-to-apples way to treat and compare machine-based and human-based solutions that target learning bottlenecks in data-to-learning-to-action processes.
But this essential nature of learning also delivers us another huge bonus: it enables rigorous quantification of the value of learning. Traditionally the value of learning has been the province of vague pronouncements of value, such as “being the right thing to do” and other gut-feel assessments. Which, of course, among other drawbacks, are highly vulnerable to the skeptical who want to cut associated budgets. But given our perspective that learning is essentially the reduction of uncertainty, we are able to draw on a relevant insight from the field of decision analysis, which informs us that decreases in uncertainty that can affect a decision have a quantifiable positive value that can be calculated.
That kind of learning that potentially affects one or more decisions we can call actionable learning, a term that has long been used with respect to human learning. But now we have given the term the more precise definition that it deserves. And that precise definition enables us to quantify the value of actionable learning that reduces uncertainty associated with limiting constraints of data-to-learning-to-action processes.
So, in summary, learning is the core feature of our business lives, yet we generally take its fundamental aspects for granted. But when we move beyond this cavalier attitude and apply a more scientific lens, the fundamental nature of learning as a process for improving predictions becomes understood. That understanding enables us to put people-based and machine-based learning on common footing and to rigorously quantify the value of learning. And that enables learning to be the cornerstone of the new approach to business performance improvement, the optimizing data-to-learning-to-action method.
Leave A Comment