Train a Model, Step by Step

You have studied the pieces: features, the training loop, error, accuracy, data splits, and failure modes. Now it is time to put them together. In this lesson you will walk through training a simple real-world classifier — not in abstract terms, but step by step, as an actual machine learning project would unfold. The model: predicting whether a fruit is an apple or an orange based on two features. Simple enough to trace every step; realistic enough to connect to the concepts you have learned.

Step 1: Define the Problem and the Output

Every machine learning project starts with a precise problem statement. Task: given measurements of a fruit, classify it as 'apple' or 'orange.' Output type: binary classification — two categories. Why machine learning? A simple rule like 'if red then apple' fails immediately — many apples are green, and blood oranges are red. The relationship between measurements and fruit type is complex enough to benefit from learning. Before collecting a single example, you decide: what exactly will the model output? A category label: 'apple' or 'orange.' This determines everything that follows — the type of loss function, the type of accuracy metric, and how you will interpret the model's predictions.

Step 2: Collect and label training data. You measure 100 fruits at a market: for each one, you record two features — weight in grams and redness on a scale from 1 to 10 (1 is fully green; 10 is deep red). You also record the true label: apple or orange. This is your labeled dataset. Step 3: Split the data. Following the principle from Lesson 7, you divide your 100 labeled fruits: 70 for training, 15 for validation, 15 for testing. The test set goes in a locked box — you will not touch it until the very end. Step 4: Choose features. Weight and redness are your features. From domain knowledge (and from a quick inspection of the data), you know these two measurements genuinely differ between apples and oranges on average. They are relevant, variable, and measurable — the three qualities of a good feature from Lesson 3.

Step 5: Train the model. You initialize the model with random parameters. Then the training loop begins: Iteration 1: The model sees a fruit with weight=182g, redness=8. It predicts 'orange' (random guess). The true label is 'apple.' Loss is high. Parameters adjust. Iteration 2: Different fruit, weight=140g, redness=3. Model predicts 'apple.' True label is 'apple.' Loss is lower. Parameters adjust slightly. This continues for all 70 training examples — one epoch. Then it loops again. After 50 epochs (3,500 total iterations), the model's parameters have settled into values that correctly classify most training examples. Step 6: Monitor with validation data. After each epoch, you check the model's accuracy on the 15 validation fruits. You notice that after epoch 30, validation accuracy stops improving and starts to slightly drop — a sign of overfitting beginning. You stop training at epoch 30 (called early stopping) and save those parameters.

Step 7: Evaluate on the test set. With training complete and all decisions made, you unlock the test set for the first time. The model classifies all 15 test fruits. Results: 13 correct, 2 wrong. Test accuracy = 13/15 = 87%. You also look at which errors the model made: both wrong predictions were apples classified as oranges (false negatives for apple). Weight and redness together were not enough information for those edge cases — perhaps small, very ripe red apples look a lot like small oranges in this feature space. Step 8: Reflect. 87% test accuracy is solid for a two-feature model on 70 training examples. The model generalizes — it was not just memorizing. To improve, you might collect more training examples, add a third feature (maybe shape roundness), or try a more complex model architecture. The evaluation guides the next iteration of development.

The Whole Picture

Let us trace every concept from this module through the fruit classifier. Machine learning (Lesson 1): we found the apple/orange pattern from data, not by writing 'if weight > 170g then orange.' Training loop (Lesson 2): the model predicted, measured loss, and adjusted parameters thousands of times. Features (Lesson 3): weight and redness were chosen for relevance and variability. Prediction (Lesson 4): the model performed binary classification — two categories, one output label per fruit. Error (Lesson 5): loss measured each wrong prediction and guided parameter adjustments. Accuracy (Lesson 6): 87% test accuracy; we also checked which type of errors the model made. Data splits (Lesson 7): 70/15/15 split; test set used exactly once. Failure modes (Lesson 8): we caught early overfitting signals via validation accuracy and used early stopping. This is what a complete machine learning project looks like — not magic, but a principled sequence of decisions, each grounded in the ideas you have been building all module.