7 perspectives on machine learning
When asked what I work on, I often have to explain what machine learning is.
I found myself using different explanations. Beyond explanations for a layperson, some perspectives are also useful for thinking about ML in general and communicating with experts:
Machine learning is about building prediction models. This is my go-to explanation of machine learning. The prediction view covers a wide range of ML applications and even LLMs (next token predictors) which many laypersons think of when hearing about AI or ML. The drawback is that it’s not perfect for unsupervised learning and prediction isn’t always the primary goal.
Machine learning is about learning patterns from data. I’ve found this view useful when working with ML practitioners and explaining my Ph.D. topic “Interpretable Machine Learning” to my parents. Many applied ML projects, especially in science, are not only focused on prediction but also on understanding, and the “patterns”-view covers both.
Machine learning is automated decision-making at scale. A view that is similar to the “prediction”-view, but emphasizes why to predict in the first place: to decide something. This view sets machine learning in a broader context that is useful when you think of ML from an economics or fairness perspective.
Machine Learning is optimization. At a technical level, machine learning is an optimization problem. This view puts the focus on math and code. A useful view if you develop new ML algorithms or work in AutoML.
Machine learning is soft computing. Instead of relying on hard-coded rules and if-else statements, machine learning allows the computer to learn “soft” patterns from data. This view is useful when you want to highlight the statistical nature of machine learning and contrast it with traditional programming.
Machine learning is compression. Prediction and compression are two sides of the same coin. This view is useful especially when explaining generative AI.
Machine learning is about algorithms producing algorithms. That’s a bit purist and meta computer science view of machine learning. It’s a useful view to highlight that ML isn’t magic.
What are views on machine learning that you found useful or insightful?
I like the "learning patterns from data" definition best. I find it a rare definition that is both simple and accurate, which is normally a difficult balance to strike for complex topics.
I also like "learn patterns from data" because it's true and relatable.
For some stakeholders, I've called it "letting the machine learn the rules of the world from the data." It speaks to tech-adverse people that know computers can be better/faster/cheaper but don't understand the outcome may be to understand the world like humans would. Want to understand rules better? Get more data and find the right way to learn the rules!
I avoid calling ML anything about prediction, because I think people overfit to the assumption of time-based predictions already. I want people to think about unique problems like unsupervised or anomaly detection that doesn't fit in their mental model of a prediction.