What Actually IS Machine Learning?
Goals, Types, and the Lifecycle of a Model.
Hi, I’m Saachi! 👋 I am anundergraduate with a passion for tech. I realized that reading about Machine Learning wasn't enough, so I started Code Train Repeat—a series where I document my journey.
The Input: Breaking the Paradigm
For years, I programmed by writing rules.
If user clicks button -> Then open window.
If temperature > 100 -> Then boil water.
This is Traditional Programming: I give the computer the Data and the Rules, and it gives me the Answer.
Machine Learning flips this upside down. I give the computer the Data and the Answer, and it figures out the Rules.
Traditional:
Input + Rules = OutputMachine Learning:
Input + Output = Rules
1. The Three Types of Learning
Machine Learning isn't one thing. It generally falls into three buckets.
A. Supervised Learning (The Teacher)
This is 90% of business ML. The data comes with an "Answer Key" (Labels).
Analogy: A teacher shows a child flashcards. "This is a cat." "This is a dog." Eventually, the child learns to identify them on their own.
The Goal: Predict the label for new data.
Examples: Predicting House Prices (Regression), Spam Detection (Classification).
B. Unsupervised Learning (The Explorer)
The data has no labels. The machine is left alone to find patterns.
Analogy: Giving a child a bucket of mixed LEGOs. They naturally sort them by color or size, even if you never told them what "Red" or "Small" means.
The Goal: Discover hidden structures.
Examples: Customer Segmentation (Clustering), Recommendation Systems ("People who bought X also bought Y").
C. Reinforcement Learning (The Gamer)
The model learns by trial and error.
Analogy: Training a dog. Good behavior = Treat (Reward). Bad behavior = No Treat (Penalty).
Examples: Teaching a robot to walk, or an AI to play Chess.
2. The Vocabulary: Speaking the Language
To work in ML, you have to stop calling them "columns" and "rows."
| Term | Definition | Simple Example |
| Features (X) | The input data used to make predictions. | Square Footage, Bedrooms, Zip Code. |
| Target / Label (y) | The answer we want to predict. | House Price. |
| Training | The process of the model "learning" the rules. | Finding the math formula that turns X into y. |
| Model | The mathematical engine (The Artifact). | The formula itself (Price = SqFt x 200). |
| Inference | Using the trained model on new data. | Estimating the price of a house that just went on the market. |
2. The Machine Learning Cycle
You don't just "train" a model once and walk away. It is a lifecycle.
A. The Golden Rule: Splitting Data
If I train my model on all my data, it will memorize the answers (Overfitting). To test if it actually learned, I need to hide some data.
The 3-Way Split:
Training Set (70%): The Study Guide. The model learns from this.
Validation Set (15%): The Practice Exam. We use this to tune settings (Hyperparameters) while training.
Test Set (15%): The Final Exam. We lock this away in a vault and only touch it once at the very end to see how the model performs in the real world.
B. The Robust Way: K-Fold Cross-Validation
Splitting data once is risky. What if my "Test Set" just happened to be really easy?
K-Fold Cross-Validation fixes this.
Split the data into K equal parts (e.g., 5 folds).
Train on 4 parts, Test on 1 part.
Repeat 5 times, rotating the test part.
Average the scores.
Analogy: Instead of taking one final exam, the student takes 5 mini-exams covering different chapters. The average score is a better measure of their knowledge.
3. The Two Main Problems: Regression vs. Classification
In Supervised Learning (where we spend most of our time), there are only two questions we usually ask:
1. "How Much?" (Regression)
The answer is a Quantity (Continuous Number).
Examples: "What will the temperature be tomorrow?", "What is the price of Bitcoin?"
2. "Which One?" (Classification)
The answer is a Category (Label).
Examples: "Is this email Spam?", "Is this tumor Benign or Malignant?", "Is this image a Cat, Dog, or Bird?"
4. The Danger Zone: Overfitting
This is the most important concept for a beginner.
Imagine a student studying for a history test.
Student A memorizes the exact dates and wording of the practice questions. (He gets 100% on the practice test).
Student B studies the concepts and causes of the events. (She gets 90% on the practice test).
On the real exam, the questions are rephrased.
Student A fails. He didn't learn history; he memorized the practice sheet. This is Overfitting.
Student B passes. She learned the general pattern, so she can handle new data.
The Golden Rule: We don't want a model that memorizes the training data. We want a model that generalizes to new data.
5. Evaluation Metrics: Keeping Score
How do we know if the model is "good"? It depends on the problem.
For Regression (Predicting Numbers)
MAE (Mean Absolute Error): "On average, my prediction is off by 500." (Easy to understand).
MSE (Mean Squared Error): Punishes large errors heavily. (Good for math).
RMSE (Root Mean Squared Error): The standard metric. It’s in the same units as your target (e.g., "Dollars").
For Classification (Predicting Categories)
Accuracy: % of correct answers. (Trap: If 99% of patients are healthy, a model that says "Healthy" for everyone is 99% accurate but useless).
Precision: "Of all the people I said had cancer, how many actually did?" (Quality of positive predictions).
Recall: "Of all the people who actually had cancer, how many did I find?" (Quantity of positive predictions).
F1-Score: The balance between Precision and Recall.
