By Paul Tibbert, CEO of GRID
“By far, the greatest danger of artificial intelligence is that people conclude too early that they understand it.” — Eliezer Yudkowsky, artificial intelligence theorist
Artificial intelligence, or AI, is perhaps one of the most discussed, broadly embraced and eagerly adopted disciplines in all of technology these days. From machine learning to sophisticated automations that run almost entirely devoid of human oversight, the business world and tech sectors are naturally enticed and intrigued by what AI has to offer our collective futures.
Yet, in the present, it can sometimes be the wrong tool for a number of modern problem sets. Understanding where to apply AI—and where to apply alternate solutions to greater effect—is perhaps the most pressing issue facing businesses today. It’s not always a question of how to tap into the power of artificial intelligence, but when.
The best way to understand the proper and appropriate application of artificial intelligence is in the context of counter-methodologies—such as the algorithm, rule setting, and human labor.
If you’re trying to solve a math problem, for example, you apply an algorithm to arrive at an exact answer—the correct one. Artificial intelligence, on the other hand, deals in probabilities. The goal with AI is not necessarily to arrive at an exact and correct answer, but rather to better understand probabilities, based on learned behaviors, and to inform a next decision based on those probabilities.
For many business problems, you do want the answer. Let’s call this the “deterministic method” of problem solving. It’s a calculation, not unlike a math equation, to determine the correct answer to a defined question.
With AI, on the other hand, what you are seeking is not so much a definitive yes or no, but rather high levels of confidence.
For many real-world business problems, what we want is actually precision. It can be misguided, then, to chase the promise and allure of artificial intelligence technology reflexively, simply because we hear so much about its future potential to be game-changing.
As Yudkowsky maintains, there’s real danger in concluding too early that we understand enough about AI to rotely apply it to all business problems of the 21st century. If, instead, you can develop and apply a deterministic method, which removes probabilities and deals in certainties, you should.
Granted, that’s not always the case.
If the real-world business case demands that the problem be solved with certainty and exactitude, a deterministic method is likely the best approach. The three most common forms of deterministic methods are tools that we’ve been applying to real-world business problems for years:
Essentially sophisticated math problems, algorithms process specific inputs, perform defined calculations applied to those inputs, and return a definitive result—the answer you seek. Algorithms, in other words, determine the unknown you are looking to solve for…not through probability, but with certainty. In my view, an algorithm is a solution every bit as sophisticated and promising as artificial intelligence, only with inferior publicity and notoriety.
Another deterministic solution is to develop a complete set of rules that can be applied to various problem inputs, sometimes unpredictable or unforeseen inputs. An example would be if/then rules: if you encounter this, the rule will determine and dictate that this, then, is the next course of action.
I sometimes call this the “brute force” method. Through the training and deployment of human labor, sometimes the best (or only) way to make accurate determinations to problem inputs is to have the human eyes, brains and hands on the job. Of course, this is where costs, time and resources tend to add up quickly, but there is sometimes no substitute for that human touch.
If any of these deterministic methods presents challenges in the problem solving process, you can always look further upstream or downstream in the lifecycle of the business or manufacturing process to identify opportunities to apply rules, algorithms or humans. In doing so, you can potentially move from “fuzzy” inputs to more concrete factual numbers by changing the point at which you’re trying to solve the problem.
The more you can eliminate uncertainty and variance early on, the less you will have to rely on confidence and probability later. Too often, I think, we over-invest in the promise of AI without applying greater scrutiny to when and where we can apply deterministic methodology to inject greater certainty and make AI tools do less guesswork and on-the-job learning later on.
Of course, not all problems are those for which you can arrive at a definitive answer using a deterministic method. In those cases, artificial intelligence is the best way to overcome the limitations of algorithms, equations, rules and human labor.
As artificial intelligence matures and refines, it certainly opens the door to solve problems we couldn’t efficiently or conceivably solve in the past. And AI holds a great deal of promise for business problems for which no deterministic solution is possible, based on the three approaches I enumerated above.
One illustrative example is language sentiment recognition software. A great many industries are enlisting “robots” to scan text, product reviews, social media conversations and more to try to understand user sentiment at scale. In this case, the cost and time it would take human labor to scan, analyze, rate and document each piece of language input in the world is unrealistic and unattainable. (Think about cataloguing the entire known universe of Yelp! reviews, for example.) Furthermore, many languages (English perhaps foremost among them) are replete with sarcasm, slang, idioms and easily misinterpreted colloquial expressions that make defining and applying a rigid set of rules nearly impossible.
However, what many are having success doing is training “machines” to learn, to pattern-recognize, and to apply probabilities based on context, previous and recognizable inputs that mimic those currently being analyzed, and with reasonable certainty, make conclusions and report decisions with acceptable levels of confidence.
Another applicable use case for AI is image recognition technology. Sophisticated machines can now capture and analyze images of machinery and parts to perform quality assessment and control at a pace and scale that far exceed human capability and capacity. This is a situation in which the use case is in need of a deterministic outcome, but for which it’s very difficult to define all expected deviations or all expected comformalities without massive amounts of human intervention and oversight. Sometimes, a person just “knows” deviation when he or she sees it, and there is no rigid set of rules that can be articulated to a machine that will reliably account for those human judgments and assessments. However, artificial intelligence can be deployed to develop a system to detect and perhaps even predict compromised quality or performance in parts, either in production or post-production, by learning what is “normal” over time and looking for any deviations from that normalcy.
In other words, AI can be “trained” to understand and recognize the conditions in image pixels that are predictive of error or risk and, through confirmation, learn how to perform this task with greater and greater certainty going forward. Ultimately, AI in this case will greatly reduce costs associated with error and eliminate waste—all in a way that is not exhaustive of expensive human labor.
The future is certainly bright for the promise of sophisticated artificial intelligence technology. And many of today’s business and technical problems call for that promise as well. Yet we shouldn’t overlook the equal promise of getting exact answers to discrete problems and questions using deterministic methodologies where we can.
Machine learning, predictability and probability, and deploying solutions that “learn on the job” are all exciting concepts and notions—as long as we are still in pursuit of precision when we can and should be.