Machine Learning: Programs Trained by Examples, Not Rules

👁 117 reads

Machine learning is a way of building programs where you don’t write rules - you show examples instead. Rather than explaining to a computer how to recognize a photo of a cat, you give it ten thousand photos labeled “cat” and ten thousand labeled “not cat,” and it derives the rule itself. You can’t read that rule like code afterward, and that’s both the source of all the power of this approach and all its problems.

Three Types Worth Distinguishing

Supervised learning. The data has answers: photos with labels, emails marked as spam, apartments with prices. The model learns to predict the answer for new cases. This is the vast majority of business applications.

Unsupervised learning. There’s no answer, just a dataset. The model looks for structure in it: groups similar customers, detects unusual transactions. You use it when you don’t know in advance what you’re looking for.

Reinforcement learning. The model acts, gets a reward or penalty, and searches for a way to maximize the reward. This is how it learns to play games, control systems, and, in simplified terms, how today’s language models are fine-tuned to respond helpfully instead of just correctly.

How This Relates to AI and ChatGPT

The relationship is nested: artificial intelligence is a broad field, machine learning is one approach within it, deep learning is a variant of machine learning based on networks with many layers, and the language models behind today’s chatbots are an application of deep learning. So when someone says “it’s AI,” they usually mean machine learning, and most often its deep variant. I wrote a broader introduction to the whole field in the article on artificial intelligence.

Where It Fails: Things Invisible in Presentations

Models inherit data biases. If there was inequality in historical credit decisions, the model will learn it and reproduce it, only faster and under the guise of numerical objectivity.

High accuracy can be an illusion. With a phenomenon occurring once in a thousand cases, a model that always answers “no” achieves 99.9% accuracy while being completely useless. That’s why for rare events you look at metrics other than accuracy alone.

The world changes, the model doesn’t. A solution trained on data from two years ago quietly loses effectiveness when customer behavior or the market changes. Nothing signals this on its own - you have to measure it.

You don’t know why. For decisions affecting people, like credit or hiring, regulations increasingly require explaining the result. This can be harder than building the model itself.

Where to Start If You Want to Understand This

Not with mathematics, but with data. Take a spreadsheet with something you know and try to predict one column based on the others. You’ll immediately see what’s most important in this field: that data quality determines the outcome more than the choice of method, and most of the work is cleaning, not modeling. Our AI dictionary will help you learn concepts, and for experiments without installing anything, try Google AI Studio or Replit.

// read moreWhat Is Artificial Intelligence and How It Really WorksGuidesApple Trained a Custom AI Model for China With Alibaba’s HelpMagazineMira Murati: OpenAI CTO, 3-Day CEO & Thinking Machines LabMagazine

Kinetyra