Machine Learning

In simple terms

Machine learning is a way of teaching a computer to do something by showing it examples instead of writing the rules by hand. You feed it data, it adjusts internal numbers to fit, and afterward it can make predictions on new data.

More detail

The three classical settings:

• Supervised learning — examples come with labels (“this image is a cat”). The model learns to predict the label.
• Unsupervised learning — no labels; find structure (clusters, low-dim embeddings).
• Reinforcement learning — an agent acts in an environment and is rewarded; learn the actions that maximise reward.

Common ingredients:

• A model (linear, decision tree, neural network, …).
• A loss function that measures how wrong the predictions are.
• An optimisation algorithm (gradient descent and friends) that nudges the model’s parameters to reduce the loss.

The dominant flavour of ML in 2026 is deep learning with neural networks, especially transformer-based language and vision models.

Why it matters

Machine learning is how modern speech recognition, recommender systems, fraud detection, search ranking, image generation, and large language models all work.

Real-world examples

• Spam filters in email.

• Recommendation engines on streaming platforms.

• The autocomplete in your text editor or IDE.

• ChatGPT / Claude / Gemini / Mistral and similar LLM products.

• Most production ML systems aren’t fancy deep learning — they’re gradient-boosted trees (XGBoost, LightGBM, CatBoost) on tabular data, which still wins most Kaggle competitions outside images/text.

Common misconceptions

• “More data always beats a better model.” Up to a point — past that, data quality and architecture matter more than raw volume.
• “ML systems are objective.” They reflect the data they were trained on, including its biases and errors.

Learn next

The dominant architecture of modern ML — see neural network.