“Is There Machine Learning That Does NOT Require Training?” is Part 1 of 2 of this article.
Is there such a thing as machine learning that doesn’t require training on sample data? The short answer is “sort of.” But to get to that explanation, we need to take several steps back.
High Quality Data and Level of Effort
At the root of the question is almost always the problem associated with the need for high quality input data and the effort required to get it. And then there is the task to train systems, which takes time and some level of skills. So even though almost everyone in business, especially technology, is enthusiastic with the “magic” of machines that learn, there is still a big effort to get to the stage where the machine is able to do something with enough precision that it is of value.
Learning Like A Human
Let’s look at how humans learn which is, after all, how we are intending our machines to behave. Humans learn through instruction and feedback. Most common methods of learning are through observation and then performing tests, getting feedback from those tests and improving. For instance, learning a language first starts with a significant amount of input (parents speaking to each other and their infant), a sponge of a brain that takes this information in (the observation), the babbling that results, and then feedback in the form of parents, relatives, or early childhood educators responding and correcting. There is no such thing as an infant that is born with the ability to speak; it is this malleable, learning approach that enables us to communicate with complex language.
With machines, the catch is that we think we want them to learn to do tasks like a human and to adapt like a human. Or, at least we think we do. The reality is that we also want those machines to be “born” with enough smarts to immediately take over tasks. This is quite a tall order to fill. There is also an analogue to that desire in the animal kingdom. All sorts of animals, from horses to giraffes, are borne and can, within a few minutes, be on their feet walking with seemingly no instruction. There are many other examples of inborn capabilities. Machines can also do the same; they can be pre-encoded to do specific tasks very well right out-of-the-box. But there are tradeoffs.
An inborn capability and the brain into which it is encoded is limited to very little adaptation. An animal’s brain into which all sorts of inborn capabilities are imbued is incapable of the wide range of learning of which humans are capable. The machine that is encoded to perform a specific task cannot do much else. So the benefit is immediate capability, and the drawback is limited capability.
Back to that answer to whether machine learning can do without training on sample data: “sort of.” The reality is that machines can be trained to perform tasks before they are “born” as products which require no training at all. They can perform these tasks very well. But they are limited to very little else. That, in the end, is not what most organizations want with machine learning.
There are advances that get us closer to the concept of “learning without learning.” However, we will save that for another day.