Every ML system, regardless of algorithm, follows the same underlying loop: represent data as numbers, define a way to measure error, and adjust the model to reduce that error — repeated until predictions are good enough.
The Four-Step Loop
| Step | What Happens |
|---|---|
| 1. Represent | Convert raw data (numbers, text, categories) into a numeric feature vector the algorithm can use. |
| 2. Predict | The model, using its current parameters, produces an output for each input. |
| 3. Measure error | A loss function compares the prediction to the true answer and produces a single number — how wrong the model was. |
| 4. Improve | An optimization algorithm (usually gradient descent) adjusts the model's internal parameters to reduce that error, then the loop repeats. |
A Concrete Walkthrough
from sklearn.linear_model import LinearRegression
import numpy as np
X = np.array([[1], [2], [3], [4]]) # study hours
y = np.array([2, 4, 6, 8]) # marks scored (toy, perfectly linear)
model = LinearRegression()
model.fit(X, y) # internally: finds slope & intercept that minimize squared error
print(model.coef_, model.intercept_) # learned parameters
print(model.predict([[5]])) # prediction for 5 study hoursExpected output: coef_ ≈ [2.0], intercept_ ≈ 0.0, and a prediction of [10.0] for 5 hours — the model learned the rule marks = 2 × hours purely from 4 examples.
What Actually Gets "Learned"
Training doesn't produce new code — it produces parameters (numbers): weights in linear/logistic regression, split thresholds in a decision tree, support vectors in an SVM. "Training a model" means searching for the parameter values that minimize error on the training data.
Practical Use Cases
This same loop underlies a spam classifier, a house-price regressor, and a recommendation engine — only the representation, the model type and the loss function change.
Common Mistakes
- Believing more complex models always "learn better" — a model can also learn the noise in the training data instead of the real pattern (see Overfitting).
- Forgetting that the loop needs a way to measure error — without a loss function there's nothing for the algorithm to optimize.
Interview Relevance
Q: "What does it mean for a model to 'learn'?" It means iteratively adjusting internal parameters to minimize a loss function computed on training data — not writing new code or rules.
Practice Question
For a spam classifier, describe what the "representation," "prediction," "error" and "improve" steps would concretely look like.
