Understanding Artificial Intelligence Processes: How Do Machines Learn?

Artificial intelligence and machine learning are the hottest corporate buzzwords right now. Even industries that seem like a non-obvious fit for the first wave of AI technology, like wine subscriptions and skin care serums, have jumped on the bandwagon to stake a claim to technological superiority. Unfortunately, many of these solutions aren’t machine learning; they’re elegant algorithms, but they don’t learn from user input or real-time results. How do machines really learn? Take a tour of the characteristics of the machine learning processes that distinguish AI from static algorithms.

AI Is Continuously Updated

Part of the learning process is the perpetual need to refine, update, and add to a knowledge base. Often, early educational inputs only approximate the real complexity of a subject; for instance, a high school physics student will likely follow a simplified curriculum that covers Newtonian mechanics but omits most other topics. As the student continues their learning, however, new information is made available to them, and their mental model of “physics” expands and changes.

Similarly, machine learning algorithms rely on a function called the optimizer, which makes changes to the algorithm’s knowledge base, or parameters, based on new information that indicates its conclusions were inaccurate, just as the physics student from the example above would eventually realize that they needed calculus-based models instead of algebraic ones to model certain phenomena correctly. There are several different types of optimizer functions; the most appropriate choice depends on what the machine is trying to learn, but the gradient descent function, which modifies parameters based on how likely those changes are to improve accuracy, is among the most popular.

AI Realizes When It’s Wrong

If the optimizer refines core assumptions, how does it know when that should happen? The loss function, another pillar of machine learning, produces the information that guides the optimizer. Put simply, the loss function describes the errors the algorithm made in its calculations. The description can be a magnitude calculation (i.e., how big was the error?) or a vector quantity (i.e., did the algorithm overestimate or underestimate?) A small absolute value for the loss function implies a high degree of accuracy.

In the physics example, the student might try to apply an algebraic model to calculate a quantity that’s better expressed by a differential, such as the rate of flow change, and discover that the answer they compute is off by a substantial margin. In this case, the loss function value might be fairly high! Like the algorithm, the student then changes their approach using a decision-making process similar to the gradient descent function; clearly, they need a new approach, but that doesn’t mean that everything they knew about physics is wrong, so they don’t want to do something completely different.

Just as for a human student, randomly selecting a new process to try probably won’t be helpful, so to increase the likelihood of getting it right, the process uses a scalar called the learning rate to make sure it only adjusts the parameters of the function a little at a time. The goal is to make adjustments big enough to matter, but not so big that helpful points on the learning curve are missed entirely.

AI Needs Minimal Input—Or None At All

The parameters the optimizer adjusts in response to the loss function’s results are the knowledge base that the machine uses as a starting point for its assumptions and calculations. Some processes draw their parameters from datasets that humans have already sorted; these are known as supervised algorithms, since a data scientist is overseeing the data before the machine receives it. Others, however, do their own sorting, scanning through piles of raw data input without human intervention. Fittingly, these are called unsupervised algorithms. If an algorithm uses machine learning, the human role in choosing outcomes is limited to this supervisory position. Algorithms where the humans adjust the machine’s behavior aren’t learning; they’re just programming.

AI is a buzzword for some businesses, but it’s also a deeply revolutionary technology, promising to one day make better decisions than the humans that created it, and despite the high-level math behind the scenes, the way machines learn is quite similar to human processes.