What You Need to Know (and Ignore) about Machine Reasoning

Recently, several technology companies have briefed me and professed to use a new type of artificial intelligence (AI) technology: machine reasoning.

If you haven’t heard the term yet, just wait. You’re going to be soon seeing it everywhere.

When I first heard the pitches, I asked if they meant machine learning, but were merely using a different term to distinguish themselves.

“No,” they assured me, “machine reasoning is completely different.”

As an industry analyst, this, of course, set me on a path to understanding the ins-and-outs of this apparently new technology, to separate the hype from reality, and to determine if this new technology was, in fact, the next big thing.

In short — and as is often the case — it’s a little bit hype and a little bit reality.

Machine reasoning is a legitimately developing segment of the broader AI sector. But it’s also incredibly nascent and will undoubtedly be subject to the frothy, over-hyped marketing treatment that has beset all things AI and digital transformation.

To help you prepare for the coming onslaught of machine reasoning hype and hyperbole, here’s what you need to know — and ignore — about it.

The Limits of Machine Learning

Machine Learning is one of the most mature, broadly applicable, and production-ready forms of AI presently available. Organizations from across the spectrum of industries are applying it to significant result across a wide range of use cases.

As you might expect, therefore, technology companies are embedding it into virtually every category of software. In fact, it’s getting difficult to find a piece of modern software that doesn’t include at least some form of the technology.

Machine learning has become table stakes.

But while it is proving to be versatile and powerful, it also comes with some substantial requirements. And they are beginning to limit its usefulness — at least as the appetite for more sophisticated intelligent applications grows.

Effectively applying machine learning requires massive amounts of data so that it can uncover the patterns that are central to its operation. Moreover, machine learning is inherently deterministic, even in its unsupervised learning form — it only works when you have a pre-determined problem, inputs, and expected outputs.

For many of the early machine learning use cases, this wasn’t a problem. Organizations knew precisely what questions they were trying to answer, and they had the data in which to find the patterns that held the answers. They just needed a machine that could expose them quickly, predictably, and automatically.

But what if you don’t have enough data to make machine learning work? Or even worse, what if you’re not even sure what questions you need to answer to solve a particular problem.

It is in these types of situations that you would turn to machine reasoning.

Machine Reasoning: What It Is and Why We Need It

In a Cognilytica article, the research firm described machine reasoning this way:

“Machine reason is the concept of giving machines the power to make connections between facts, observations, and all the magical things that we can train machines to do with machine learning.”

In effect, machine reasoning is about making a machine approach information the same way humans do: by understanding its essence so that it can use that understanding to process and comprehend data that would otherwise have no context. Or, as the same Cognilytica article put it, “to functionally use that information for higher ends, or apply learning from one domain to another without human involvement.”

From a very young age, children apply both inductive and deductive reasoning in just this way as they learn. For instance, tell a small child that all birds have feathers, and then separately explain that the parakeet he’s looking at is a bird, and he will rightly reason that all parakeets have feathers.

A machine learning algorithm, however, will not. It is unable to connect otherwise unrelated datasets to come to such a conclusion.

The usefulness of a machine reasoning system, therefore, becomes readily apparent in any situation in which you have only limited data or in which that data is highly volatile such that you cannot build a reliable machine learning model based upon it.

Likewise, problems that have no directly correlated dataset are problematic for machine learning algorithms as it’s impossible to detect patterns in data when you don’t know which data to analyze.

A human, of course, would use reasoning to deduce the potential sources of the problem, and then examine the relevant datasets, identifying relationships between them as she went until a possible answer emerged. She would then test that answer and start the process over again if it didn’t fit.

While this approach would be impossible for a machine learning model that requires a defined dataset, it’s precisely the type of problem that machine reasoning can solve.

The only problem is that it doesn’t really exist. Yet.

Beware Machine Reasoning Hype

Despite that fact that we’re starting to see machine reasoning pop-up in messaging and marketing materials, the reality is that true machine reasoning is not yet here.

Being able to infer context requires conceptual domain awareness — the ability to intuit what may be relevant and what may not be in any given situation — that is proving elusive to codify in machine reasoning models, except in very narrow-banded use cases.

Likewise, while the computational demands of machine learning are still pushing the boundaries of currently available, affordable, and scalable resources, machine reasoning is “easily one order or more of complexity beyond machine learning,” according to Cognilytica.

Still, that doesn’t mean that the tech companies that are beginning to use the machine reasoning moniker are just blowing hot air. These companies are, in fact, applying elements of machine reasoning approaches to address the machine learning gaps.

The challenge, of course, is that to accomplish this feat, they must apply these approaches in very narrow and targeted use cases — those in which they can significantly narrow and define the universe of potential relationships and contextual domains.

These early applications are showing promise in their ability to predict outcomes or events with limited and loosely-related datasets, but under the covers, machine learning algorithms are still doing much of the heavy lifting.

Don’t Get Swept Up

I expect that you will hear a lot about machine reasoning in the coming months and years.

Tech company marketers are always on the hunt for anything that will help differentiate them in a crowded and noisy market. And the allure of a new technology that can solve new problems and work with limited data will be just too juicy for most of them to pass up.

And, as is almost always the case, there will be elements of truth to the pitch. They will have new, patented approaches that help solve specific problems better, faster, and less expensively.

So what’s a progressive enterprise leader to do?

If a particular machine reasoning-based solution helps you solve a pressing business problem with which you’ve been struggling, then great! Swipe right, make a match, and go for it.

Otherwise, take care not to get caught up in the next great wave of hype. True machine reasoning is likely some ways off and — trust me on this one — you’ll know it when it gets here.

Until then, remain steadfastly focused on one thing: how to harness technology to delight your customers, partners, and employees, and create a competitive advantage in the process. Whether that comes in the form of a machine reasoning solution or not is beside the point.