人工智能：我們是在教機器學習，還是隻教它們辨認模式？

Artificial Intelligence: Are We Teaching Machines to Learn or Just Recognize Patterns?

While exploring Machine Learning, I started questioning something fundamental:

Are we really teaching machines to learn, or are we mainly teaching them to recognize patterns from examples?

I believe there is an important distinction between learning individual examples and understanding the fundamental principles behind them.

Learning from examples vs. understanding principles

An example is a specific situation from which we can learn something. A principle, however, is a fundamental truth, relationship, or structure that helps explain why something works.

These principles are particularly valuable because they can be reused across different contexts. Once we understand the underlying relationship, we can apply that knowledge to situations we have never encountered before.

This is one of the most powerful aspects of human learning: we don't necessarily need to memorize every possible situation. We can understand a principle and use it to reason about new problems.

Consider two developers preparing for algorithmic coding interviews:

Developer A solves 300 coding problems, focusing mainly on practicing individual problems and remembering their solutions.
Developer B solves 65 carefully selected problems, but focuses on understanding the underlying principles and algorithmic patterns, such as Hash Maps, Two Pointers, Sliding Window, Binary Search, Recursion, and Dynamic Programming.

Who is more likely to solve an unfamiliar problem?

Developer B may have an advantage because they are developing reusable problem-solving structures rather than relying primarily on familiarity with previously solved examples.

The point is not that solving 300 problems is wrong, or that 65 is a magic number. Practice is essential. The difference is what each person extracts from that practice.

One focuses on accumulating solutions. The other focuses on understanding the principles that can generate solutions across different problems.

This distinction is what interests me most about the future of Artificial Intelligence.

What does a Machine Learning model actually do?

Let's use a simple everyday analogy: imagine cooking a pot of food for the first time.

You taste it and realize it needs more salt. You adjust the seasoning, taste it again, and make another adjustment.

Each attempt gives you feedback that helps you get closer to the result you want.

This is learning through trial and error. However, it is not necessarily blind trial and error. The feedback from each attempt guides the next adjustment.

Many Machine Learning models follow a related mechanism. They make predictions, compare them with actual outcomes, measure the error, and use that information to adjust their internal parameters.

For example, in Linear Regression, Gradient Descent uses mathematical information about the direction of the error to guide parameter updates.

This process can be effective. However, repeatedly making adjustments without understanding the underlying relationships can become inefficient.

In human learning, trial and error is a natural mechanism, but it is not the only one. We can also learn by identifying fundamental principles, understanding relationships, and transferring that knowledge to new situations.

This brings me back to my original question:

Can we develop AI systems that go beyond learning statistical patterns from examples and become better at identifying, representing, and reusing fundamental principles?

From theory to practice

To understand one of these learning mechanisms beyond the theory, I implemented a Linear Regression algorithm from scratch, without using Machine Learning libraries.

This allowed me to explore the mathematics behind the process, including model parameters, the cost function, Mean Squared Error (MSE), and Gradient Descent.

I started with the mathematical formulation and progressively translated it into code. This experience helped me understand how a model makes predictions, measures its errors, and adjusts its parameters to improve.

More than implementing an algorithm, this project changed how I look at Machine Learning and reinforced my interest in understanding the mechanisms behind learning itself.

Perhaps the next major challenge in Artificial Intelligence is not simply learning from more data, but learning more from what the data represents — identifying the principles behind examples and making that knowledge reusable across different contexts.

That is one of the questions I want to keep exploring.

github: my code on github