Machine Learning Explained: Types, Uses, and What It Can't Do

Machine learning is the technology behind facial recognition, spam filters, medical diagnosis tools, recommendation systems, and large language models like ChatGPT. The term is used constantly, often interchangeably with “AI” — but they’re not the same thing, and the distinction matters.

Machine Learning vs. AI

Artificial Intelligence is the broad goal: building systems that can perform tasks that would normally require human intelligence.

Machine Learning is the main approach used to achieve that goal today: systems that learn from data rather than being explicitly programmed with rules.

Traditional programming: programmer writes rules → computer follows rules → output.

Machine learning: programmer provides data + desired outputs → algorithm learns rules → output.

The power of ML: instead of trying to write explicit rules for recognizing a face (which is nearly impossible), you show the system millions of labeled faces and let it figure out the rules itself.

The Three Main Types

Supervised Learning The most common type. You provide labeled training data: thousands of emails labeled “spam” or “not spam,” medical images labeled “tumor” or “no tumor,” loan applications labeled “defaulted” or “repaid.”

The algorithm learns to map inputs to outputs. After training, it can classify new examples it’s never seen.

Applications: spam filters, image classification, fraud detection, medical diagnosis, credit scoring.

Unsupervised Learning No labels. The algorithm looks for patterns and structure in the data itself.

Clustering: group similar items together (customer segmentation, document categorization). Dimensionality reduction: find the most important features in complex data (useful for visualization and preprocessing). Anomaly detection: identify unusual patterns (credit card fraud, network intrusion detection).

Reinforcement Learning An agent learns by interacting with an environment, receiving rewards for good actions and penalties for bad ones — similar to how you train a dog.

Applications: game-playing AI (AlphaGo defeated the world Go champion using RL), robot control, optimizing trading strategies, recommendation systems.

Key Concepts

Training data: The examples the model learns from. Quality and quantity of training data largely determines model performance.

Features: The input variables the model uses to make predictions. For a house price model: square footage, location, number of bedrooms, age.

Model: The mathematical function that maps inputs to outputs, learned during training.

Overfitting: When a model memorizes training data instead of learning generalizable patterns. Performs well on training data, poorly on new data.

Underfitting: When a model is too simple to capture the patterns in the data. Poor performance everywhere.

Validation / Test set: Data held back from training, used to evaluate how well the model generalizes to new examples.

Deep Learning: The Current Driver of Progress

Deep learning is a type of machine learning that uses neural networks with many layers (hence “deep”). It’s responsible for the dramatic AI progress of the past decade.

Deep learning powers:

• Large language models (GPT-4, Claude, Gemini)
• Image generation (Midjourney, DALL-E)
• Speech recognition (Siri, Alexa)
• Real-time translation (Google Translate)
• Medical image analysis
• Self-driving vehicle perception

Why did it take off? Three things aligned: large datasets (the internet), powerful GPU computing, and the Transformer architecture.

Where ML Struggles

Despite impressive capabilities, machine learning has real limitations:

Data hungry: Most ML approaches need large datasets. Getting enough labeled data is often the bottleneck in real applications.

Brittle generalization: Models can fail unexpectedly on inputs that are slightly outside their training distribution. An image classifier trained on certain lighting conditions may fail in different lighting.

No causal reasoning: ML models learn correlations, not causes. A model might learn that ice cream sales correlate with drowning deaths (both peak in summer) — without understanding that neither causes the other.

Opacity: Complex models (especially deep neural networks) are difficult to interpret. You know they work, but often not why — a problem for high-stakes decisions like loan approvals or medical diagnoses.

Data bias: Models learn the patterns in training data, including human biases. Facial recognition systems have shown lower accuracy on darker-skinned faces due to imbalanced training data.

Computational cost: Training large models requires enormous computing resources — millions of dollars in GPU compute for frontier models.

Practical ML Today

Machine learning in production is less glamorous than headlines suggest. Most deployed ML systems are:

• Linear or logistic regression models for straightforward prediction tasks
• Gradient boosted trees (XGBoost, LightGBM) for tabular data
• Pretrained models fine-tuned for specific tasks, rather than trained from scratch
• Simple rule-based systems for well-defined problems, with ML filling the gaps

The dramatic models you read about (GPT-4, AlphaFold) represent the frontier. Most working ML is simpler and more pragmatic — solving specific, well-defined problems with enough data.

---

Machine learning is pattern recognition at scale: algorithms that discover structure in data without being explicitly programmed. Powerful for the right problems, brittle in unexpected ways, and fundamentally dependent on the quality and quantity of the data it learns from.