Artificial Intelligence (AI) can be quite a challenging topic to truly comprehend, especially for business managers, entrepreneurs and investors that lack a deep academic background in the field. They may instinctively sense the massive potential of AI — all the science fiction movies and TV shows that Hollywood churns out probably plays a part in this — but they are often left wondering, how should I think about AI? How does AI actually work?
The follow article addresses this gap by presenting two broad and fairly dominant intuitions of AI — cognitive and statistical. Despite the relative fragmentation of the field and varied backgrounds of AI practitioners, the cognitive and statistical intuitions seem to reflect the ways of approaching AI today. If you can grasp one or both of these intuitions, then you will be better positioned to meaningfully participate in discussions around AI as a business stakeholder, as well as build and invest in AI opportunities.
Think about the last time you had to study for a test with multiple choice questions. Fig. 1 shows a very simple example of such a question.
If you were lucky, you may have had access to a question bank of past tests to practice on. If you were like me, you probably worked through the questions several times, hopefully getting better and better with each try.
This is also essentially the cognitive intuition of how AI works. It emphasizes the notion of learning and drawing parallels to human cognition. Similar to a student working through sample questions from past tests, the AI uses past data to build an increasingly accurate understanding (or model) of the real world. In particular, both you and the AI do the following:
- Train with data: The more multiple choice problems you work through, the better you get at the task. Of course, by doing this you typically only getting better at the specific task at hand. For instance, if you have been studying for a biology test, you are unlikely to improve your performance on a history test. The same goes for AI.
- Learn by trial and error: You want to maximize rewards (by getting answers right and acing the test) and avoid punishment (minus points). You can also think of AI learning like this; it is rewarded for correct predictions and punished for wrong ones. Over time, the AI adapts its behavior to cut out mistakes and maximize rewards.
- Simplify reality: To make the learning task a little easier, you might use simple heuristics, shortcuts or other tricks to improve your accuracy. For example, you could group similar concepts into clusters, or use mnemonics to help you recall the subject matter better; all of these things can help you get to the correct answer without having to work through the question in detail. Similarly, AI also relies on models that are simplifications of the real world to respond quickly to the environment. Think of self-driving cars maneuvering around obstacles on the road, or even your phone using facial recognition to unlock the screen — the detection does not need to be perfect, just good enough to approximate reality.
Back in high school, and later at university, I remember drawing best-fit lines through points plotted on a scatter graph. Fig. 2 illustrates a simple example in which the objective is evidently to draw a straight line that is as close as possible to as many of the points on the graph as possible.
With a pretty basic understanding of best-fit lines, you are already in a position to understand how AI works. This is because AI builds on top of statistical concepts to develop intelligent ways of thinking about the world — and that is also the basis for the statistical intuition of AI. In fact, just from the graph in Fig. 2, we can begin to understand some concepts that are fundamental to AI:
- Uncovering relationships: AI is all about uncovering oftentimes very complex relationships between different variables in the data. The best-fit line in Fig. 2, for example, describes a possible (and pretty simple) linear relationship between the variables x and y. Relationships can also be non-linear (e.g., quadratic, sinusoidal, etc.) and span more than two dimensions. If the relationship can be represented using math, then it is probably fair game for AI.
- Interpolating and extrapolating: This has to do with filling the gaps in the data, such as predicting the value of y for values of x (and vice versa) that are not already plotted on the graph. If the missing data is within the range of our current data sample, then we talk about interpolation. If the missing or unknown data is outside the range of our data sample, then we are concerned with extrapolation.
- Detect anomalies: Several high-impact use cases of AI have to do with detecting outliers or anomalies in data. In manufacturing, for instance, production processes rely on the stability and consistency of equipment; AI is increasingly used to facilitate “predictive maintenance” (e.g., by monitoring and acting on the health status of equipment) to preempt degradation in the equipment performance and thus reduce downtimes along the production line. Spotting anomalies and thinking about what might have caused them (e.g., via a root cause analysis) are therefore key issues for AI.
- Understanding issues of underfitting and overfitting: The relationship between x and y should fit known as well as unknown/new data. For example, the line in Fig. 2 might fit the current data sample, but what if the relationship between x and y changes as x becomes large? For all we know, the slope of the best-fit line could stay the same as in our current data sample, but the slope could also decrease and perhaps even exhibit a downward trend. Underfitting means that the statistical relationship uncovered by the AI only roughly fits the data (and that it probably does not fit the data well enough), whereas overfitting means that the relationship may be fitting the available data sample too well (and may be way off for data that we have not yet seen). Achieving the right balance here is a key challenge in AI.
It can be quite difficult to wrap ones head around AI. Despite being around for nearly a century now, its inherent complexity can make AI feel like a nascent field that is still more science fiction than reality. Business managers, entrepreneurs and investors can instinctively sense that AI is — or will soon become — a pretty big deal, but the technology can still often feel like a black box.
To open up this black box and shed more light on how AI actually works, it can help to think in terms of the two dominant intuitions of AI presented above. The cognitive intuition comes at AI from the perspective of human cognition and learning; using our own cognition as an analogy can make it easier to grasp how AI works. Meanwhile, the statistical intuition looks at AI through the lens of mathematical relationships; this intuition is all about uncovering, describing and optimizing the oftentimes complex and seemingly hidden relationships between the outcome variable of interest (the y) and possible predictors (the x’s). Taken together, the cognitive and statistical intuitions provide a fairly holistic picture of how AI works under the hood.
Finally, having understanding of these two dominant intuitions of AI is also highly relevant for those of you actively recruiting for data scientists and other AI professionals today. You will notice that most candidates primarily make sense of AI through either the cognitive or statistical viewpoint. For example, candidates with backgrounds in economics and mathematics (and, of course, statistics) tend to transition to AI from the world of statistics; as a case in point, one former economist that I interviewed for a data science role said that AI under the hood is basically just “statistics on steroids”. At the same time, there are candidates with backgrounds in computer science or mechanical engineering, that may have experience in automation and robotics; for such candidates, the jump to AI tends to be a shift in thinking from rule-based, deterministic machines to more “intelligent”, non-deterministic machines that can interact with the environment in an adaptive manner. Arguably, recruiting for both the cognitive and statistical intuitions can lead to more well-rounded and effective AI teams.
All Rights Reserved for Chinmay K