What Makes a Machine Learn? Understanding the Journey from Data to Decisions | by Prerna Waghray

1️⃣ Introduction: Learning Without Knowing

When we talk about “Machine Learning,” the word “learning” can be misleading.
Machines don’t “understand” in the human sense — they optimise, adapt, and most importantly, generalise. Rather than attaining understanding, they find mathematical patterns and improve performance based on past data.

2️⃣ The Heart of Learning: From Data to Decision

At its simplest, a machine learning algorithm is a function that links an input to an output:

Y^=f(X)

where X is the input data and Y^ (Y-hat) is the model’s prediction.
The model learns by comparing its prediction Y^ to the actual value Y and minimising the difference (the error) between them — a process called training. Many algorithms, such as linear regression or neural networks, use gradient descent and other optimization methods to reduce this error over time.

3️⃣ The Learning Loop

All ML models — from simple regression to deep neural networks — follow a fundamental cycle:

Collect and preprocess data: Gather relevant data and clean it to ensure accuracy.
Train a model: Discover patterns and relationships using chosen algorithms.
Evaluate performance: Measure prediction accuracy or classification results against real outcomes.
Adjust: Fine-tune model parameters to improve generalization and avoid overfitting.

This loop mimics curiosity and adaptation — trying, testing, failing, and improving in each cycle.

4️⃣ Types of Learning

Supervised Learning: Guided by labeled examples (e.g., predicting house prices from historical sales data or classifying spam emails).
Unsupervised Learning: Finds patterns without explicit labels (e.g., grouping customers based on shopping habits or discovering clusters in data).
Reinforcement Learning: Learns through trial and error by receiving rewards or penalties (e.g., self-driving cars learning to navigate, or an AI playing chess, robotic vacuum).

5️⃣ The Power of Generalisation

A good model doesn’t just fit the data it’s seen — it accurately predicts outcomes on new, unseen data. This ability, called generalisation, is the real core of “intelligence” in machine learning.

6️⃣ Bias, Variance, and the Balance

High bias: The model is too simplistic; it misses important relationships (underfitting).
High variance: The model is too complex; it memorises noise as if it’s meaningful (overfitting).
Balance: The key to a successful learning system is striking the right balance, enabling it to generalise well.

7️⃣ Visual Concept

A simple flow diagram explains the process well
Data → Preprocessing → Model → Prediction → Error Calculation → Update → Improved Model
Each arrow represents a step in learning, showing how every stage, from raw data to updates, drives progress.