Artificial Intelligence (AI) is everywhere: big tech, small tech, Hollywood, politics, elections, financial markets, sports, the cloud, your pocket, education, farming, pharmaceuticals, bio-engineering, geological exploration, physical modeling… In fact, it’s reasonable to assume that if you aren’t already using AI in your current job, you will be in the next 2–7 years. And if not, something’s wrong.
But AI comes in many flavors. It takes on many forms. And we often talk about one flavor, when what’s really happening is a different flavor, or a mix of different flavors.
I had the privilege and pleasure of working in tech (mostly research) for 30 years. In hindsight, most of that work was about developing AI. Any success I had there was generally the result of making complicated systems simple and direct, and making them work well. While it’s dangerous to over-simplify complicated systems, thinking about these complicated systems in simplified forms gives us an internal model that can help us decide how we want to to use AI, and what we might expect from it.
By analogy, few of us understand the complexities of a modern car, but we have a sufficient (simplified) model of what it is, and how it works, that allows us to use a car well.
In that simplified context, there are 3 basic flavors of AI: (1) expert systems, or rule-based systems; (2) data-driven structured models; and (3) unstructured, or end-to-end data-driven models.
Most of the interesting “planet scale” AI systems are a hybrid mix of all 3. Just like nature is loath to get rid of things that work well (DNA, symmetry, eyes, photosynthesis, survival instincts), AI is the product of evolution, and incremental improvements are often made by building on top of what’s already there. That said, it’s also reasonable to argue that over the last 40 years, expert systems were generally replaced by data-driven structured models, and then those systems were generally replaced by end-to-end unstructured models.
My definition of AI in this article includes systems where machines do something that requires judgement about natural data that we might typically associate with human judgement. I’m not talking about trying to build a system that tries to pass as human — the original Turing test. That’s a creative test, but we already have systems that are much better than humans at many existing tasks, and pushing the limits of those systems, pushing their Artificial Intelligence, and exploiting them well is already extremely important for society.
To describe those 3 generations of systems, we’ll use a trivial task: gender identification. The goal of our toy AI systems will be to identify the gender of a person. As a culture we’re learning that (just like the Kinsey report showed us for sexual preferences) gender itself is more continuous, and less binary, than most of us imagine. But for these toy systems, we’ll start by only considering binary gender.
A gender ID system is a completely fictitious and ridiculous system that would never be built. And if we can get past that, it’s also a helpful example to think through the different approaches to AI.
Expert systems model the debate culture of academia that probably started with Socrates. If this, then that. Logic. Let’s make a system that asks the right questions, and we’ll make rules to map the answers to those questions to the decisions of the system. That’s the logical way to do these things...
You have to be careful with expert systems.
