What Does It Mean When We Say AI "Learns"?
In the previous lessons, we have repeatedly used the phrase "the model learns from data." But what does that actually mean?
A computer does not learn in exactly the same way a human learns. When we say that a machine learning model learns, we mean that the model adjusts internal mathematical values so that it becomes better at a particular task.
The basic process looks like this:
Data → Training → Model → Prediction
To understand this process, we need to understand a few important ideas: training data, models, parameters, training, and inference.
What Is Training Data?
Training data is the collection of examples that a machine learning system uses during training.
The type of data depends on the problem being solved. It could contain images, text, audio, numbers, video, sensor measurements, or business transactions.
For example, imagine building a system that identifies whether an email is spam. The training data might contain thousands of emails that have already been classified as either spam or not spam.
The model can examine those examples and learn patterns that help it classify future emails.
Examples and Labels
Some machine learning problems use examples that include a label. A label tells the model what the correct answer is for that example.
For our spam example:
- The email is the input.
- "Spam" or "Not Spam" is the label.
Another example would be predicting house prices. The information about the house might include its size, location, and number of bedrooms, while the selling price is the value the model is trying to predict.
Not every type of machine learning uses labels in this way. Later in the course, we will look at different approaches to machine learning.
What Is a Model?
A model is the learned mathematical system that uses patterns in data to produce an output.
Before training, the model does not yet have the useful patterns needed for the task. During training, the model's internal values are adjusted based on the examples it receives.
After training, the resulting model can be used to make predictions on new data.
A useful simplified way to think about it is:
Training data + learning process → trained model
What Are Parameters?
Modern machine learning models contain internal numerical values called parameters.
These parameters influence how the model processes information and produces its output.
During training, the learning process adjusts these parameters so that the model becomes better at the task.
You do not need to understand the mathematics behind every parameter to understand the basic concept.
A simple analogy is tuning a large collection of controls. The model has many internal controls, and training gradually adjusts them so that the model produces better results.
Large modern AI models can contain enormous numbers of parameters. The important beginner-level idea is simply: parameters are internal values that the training process adjusts.
What Happens During Training?
Training is the process through which the model learns useful patterns from examples.
A simplified training process looks like this:
- The model receives an example from the training data.
- The model produces a prediction.
- The prediction is compared with the expected result when one is available.
- An error or loss is calculated.
- The training process adjusts the model's parameters.
- The process is repeated many times across many examples.
The goal is not simply to make one prediction correct. The goal is to adjust the model so that it performs well across many examples.
What Is Loss or Error?
A machine learning system needs some way to measure how well its prediction matches the desired result.
This measurement is commonly represented using a loss function.
If the model's prediction is poor, the loss can be larger. If the prediction is closer to the desired result, the loss can be smaller, depending on the particular problem and loss function.
Training uses this information to determine how the model's parameters should be adjusted.
You can think of loss as a signal that tells the learning process: "How far away was the model's answer from what we wanted?"
A Simple Example
Imagine a model that predicts house prices. Suppose the actual selling price of a particular house is ₹50 lakh.
The model initially predicts ₹35 lakh. That prediction is not very accurate.
The training process measures the error and uses it as part of the process of adjusting the model's internal parameters.
After seeing many examples, the model may become better at estimating house prices.
The model is not being given a rule such as: "A three-bedroom house costs exactly ₹50 lakh." Instead, it is learning relationships between many variables and the target value.
What Is Inference?
Once a model has been trained, we can use it to process new data. This stage is commonly called inference.
For example:
Training: The model learns from many examples of spam and non-spam emails.
Inference: The trained model receives a new email and predicts whether it is likely to be spam.
So a useful distinction is:
Training = learning from examples.
Inference = using the trained model on new inputs.
Training Does Not Mean Memorizing Everything
It is tempting to imagine that a model simply stores every training example in its memory. That is not the basic goal of machine learning.
The goal is for the model to learn patterns and relationships that allow it to perform well on new examples.
For example, a model trained to recognize cats should ideally recognize a new photograph of a cat that was not part of its training data.
This ability to perform well on new, unseen examples is an important part of machine learning.
What Is Generalization?
The ability of a model to perform well on new data is called generalization.
A model that simply performs extremely well on its training examples but performs poorly on new examples has not learned the underlying patterns as effectively as we would like.
This leads to an important problem called overfitting.
What Is Overfitting?
Overfitting happens when a model becomes too closely adapted to its training data and does not perform well on new data.
Imagine a student who memorizes the answers to a practice test but does not understand the underlying subject. The student might score perfectly on the practice questions but struggle when the questions are changed.
A machine learning model can have a similar problem. It may perform very well on the examples it has seen but fail to generalize to new examples.
Machine learning techniques are designed in part to help models learn useful patterns rather than simply memorize the training examples.
Why Data Quality Matters
The quality of the training data has a major influence on the quality of the resulting model.
If the training data contains errors, missing information, unrepresentative examples, or unwanted bias, the model can learn those problems.
This leads to a simple but important principle:
Better data can lead to better AI systems.
But more data is not automatically better. The data also needs to be relevant, useful, representative, and appropriate for the task.
Why Training Data Needs to Be Representative
Imagine training an image recognition system using photographs taken only in bright daylight. The system might perform well on similar images but struggle with images taken at night.
The problem is not necessarily that the model is incapable of recognizing the object. The training data did not adequately represent the conditions under which the model is expected to operate.
This is one reason why designing and preparing training data is such an important part of machine learning.
Training Data, Validation Data, and Test Data
In many machine learning projects, the available data is divided into different groups.
Training data is used to train the model.
Validation data can be used during development to help evaluate and improve choices about the model.
Test data is used to evaluate how well the final model performs on data that was kept separate from the training process.
The exact way data is divided depends on the problem, but the general idea is important: we want to evaluate whether the model can work well beyond the examples it learned from.
Human Involvement Is Still Important
Machine learning can automate much of the pattern-learning process, but humans remain important.
People decide what problem to solve, collect and prepare data, choose appropriate methods, evaluate the results, and determine how the model should be used.
People also need to consider issues such as privacy, fairness, security, safety, and the consequences of incorrect predictions.
The Big Picture
We can now describe the basic machine learning process more clearly:
Data → Model → Prediction → Measure Error → Adjust Parameters → Repeat
During training, the model repeatedly processes examples and adjusts its internal parameters so that its performance improves.
After training, the model can be used for inference on new data.
The real goal is not to memorize the training examples. The goal is to learn useful patterns that generalize to new situations.
Understanding this basic process gives us a much stronger foundation for everything that follows in AI.