You might be detecting a theme in the last few posts: AI isn’t always the answer. Sometimes it’s better to use just humans, sometimes humans in the loop with a non-AI solution. Here are a few other alternatives you should be sure to consider.
The many flavors of machine learning
At the moment, most enterprise AI systems are built using machine learning (ML), along with its modern “offspring”, deep learning.
But machine learning can be accomplished using a number of techniques, including decision trees, regression trees (RTs), random forest (RF), support vector machines (SVMs), Bayesian networks, clustering, and many more, each with its own strengths and weaknesses in a given application. All of these methods have been called “Machine Learning” at one time or another.
Other paths to AI
Of course, machine Learning and AI aren’t synonymous. Other ways to achieve what has from time to time been called “artificial intelligence” include:
- A myriad of probabilistic techniques,
- Kernel methods (decision trees/random forest, Gaussians, PCA),
- A slew of reinforcement learning techniques that use symbolic AI rather than neural networks, such as artificial reasoning a.k.a. “good old fashioned AI” (GOFAI),
- Path planning and intelligent control systems methods that are labeled “classical AI” (not the same as GOFAI),
- Artificial life (swarms, cellular automata, and other simulations inspired by natural processes),
- Agent-based modeling (this is what gamers call “AI”)
- Chaos modeling systems,
- Topological search systems,
- and more. You may have your own favorite here
Smart AI people know how to use the right tool for the job, and their toolbox includes more than a few methods. They know that the analytic techniques and machine learning algorithms that will meet the core business and cultural needs of manufacturing, energy, and transportation firms are not necessarily going to be the same as the ones used by advertisers and marketers to place banner ads or recommend which hair conditioner is best for you based on the PH of your favorite shampoo.
On the other hand, you can go far into “solutionism”: always wanting to apply the sexiest new method, when a more proven and stable method is a better fit. Stability, proven scale, production-readiness, Technology Readiness Level (TRL) – all of these are critical attributes of your problem-to-algorithm fit analysis.
Success not only requires doing the hard work to understand which tool is right not only for where the customer is today, but also where they want (or will be forced to) take their business in the future. It requires a toolbox that is big enough to do the job, but avoids obsession with the latest and greatest.
Here’s the kind of thinking I’m talking about. This 2015 paper looked at a use case for ML applied to identifying the location of and predicting accessibility (that is, cost to extract) of mineral deposits: Its thinking about the right ML algorithm was that the, “…results of applying…[ML]…algorithms to epithermal Au prospectivity mapping of the Rodalquilar district, Spain, indicate that the RF outperformed the other MLA algorithms (ANNs, RTs and SVMs). The RF algorithm showed higher stability and robustness with varying training parameters and better success rates and ROC analysis results. On the other hand, all MLA algorithms can be used when ore deposit evidences are scarce.” In other words, multiple characteristics of the algo were taken into account.
More generally, you and your customer must establish what conditions they expect to operate in and what results are desired. This context provides critical information for determining which ML algorithm will provide acceptable results at an acceptable ROI.