Artificial Un-Intelligence

Despite what you may have heard, Artificial Intelligence is not a one-size-fits-all solution. Machine Learning, AI, regressions, supervised, unsupervised, forecasting, classification, similarity matching, and even good-old brute-force models all have their place in data science; why then do people jump to the conclusion that AI is best?  There is a lot of hype going around, even calling these past few years ‘the revitalization of AI’.  And it’s mostly true: self-driving cars, financial models that can reverse-engineer the formerly impossible, and translation software that is getting eerily close to the real-deal.  But don’t go thinking that AI can solve every problem. It can’t. And even if it can, it may not be the right solution.

To understand why I’m saying this, let’s revisit what a ‘model’ actually is: it’s a digital representation of a real phenomenon. It provides you with a very good framework from which you can distil the real solution. But that’s it: it’s an estimate.  A ‘good’ model then, is one that accurately and simply replicates the results. But even still, it in and of itself, a model is not the true relationship unless you can reject all other models. This last part is an important stipulation because to reject all other model-solutions can be very difficult, especially in the case of neural networks and artificial intelligence.

Artificial Intelligence models, particularly deep-learning models involving many layers of neural-networks, are notoriously difficult to build, train, and distil information from once complete.  Yes, there are new packages that streamline this process (ex: TensorFlow), but the fact remains that AI is by no means ‘low-hanging-fruit’.  Moreover, it takes expert knowledge to apply AI in your specific use-case and even more, knowledge to understand whether it’s actually the optimal solution.

So where does this leave us as data scientists, statisticians, and business professionals?  If it’s all guesswork and estimations, how then can we truly reconcile each of these methods into properly-informed business decisions?  Often times, professionals massage their data into AI-templates and pipelines, building unnecessarily complex models when perhaps a simple regression would do. Moreover, even if this unnatural coercion of data provides a solution, it may be the answer to an entirely different problem—and you might never know that your ‘good’ results could be useless. In total, my suggestion is to find yourself “A Keymaker”…

“The Matrix Reloaded”, the second film in the Matrix series, may not have been as good as the first, but it portrays one of my favorite characters of all time: The Keymaker.  In the film, the Keymaker is a minor character who has the outstanding foresight to make any key for any lock.  Whether it be a door or a motorcycle, the Keymaker always seems to have the solution.  Such is the life of a consultant; it’s our job to understand not just the problem at hand but to architect a solution that meets the needs of all your business goals.  It’s the consultant’s job to anticipate our clients’ needs, and know which model will fit for which problem.

So when deciding if Artificial Intelligence is right for your use-case, it is important to always start by stating your problem in the most efficient way possible.  If not, you could rush into an AI model, spend three months building a pipeline and wasting the time of your developers and engineers, only to find that you needed an entirely different class of model or regression. I’ll end with two quotes from “The Keymaker”, I hope that you as the reader can appreciate their relation to this discussion on Artificial Intelligence and data science models in general: One, “We do only what we’re meant to do” and two, “Another way. [There is] Always another way”.