What is machine learning?

Definition

What is machine learning?

Machine learning is a branch of artificial intelligence in which algorithms learn from data and make predictions – without developers programming every rule individually.

Difference from classic software

A conventional program follows fixed rules that a human specifies. An ML model recognises patterns in data and derives its own rules from them.

Example: spam filter

Instead of maintaining a list of banned words, an ML-based spam filter analyses thousands of emails and learns for itself which characteristics indicate spam. The more emails it processes, the more accurate it becomes.

How it works

How does machine learning work?

An ML system goes through four phases – collecting data, training the model, evaluating results and using the finished model in everyday operation.

Collect and prepare data

Everything starts with data. Developers collect raw data, clean up errors and split the dataset into training and test data. The quality of the data determines the quality of the model.

Train the model

The algorithm processes the training data and adjusts its parameters step by step. Goal: the predictions should match the actual results as closely as possible.

Evaluate results

The test data check whether the model also works correctly with new inputs – or whether it has only memorised the training data. Experts call this problem overfitting.

Deploy the model (inference)

The trained model now processes new, unknown data and delivers predictions in real time. This step is called inference – the model applies what it has learned to practice.

Learning methods

What types of machine learning are there?

Machine learning is divided into three learning methods – supervised learning, unsupervised learning and reinforcement learning.

Supervised learning

The model learns from labelled data: every input comes with the correct answer. The algorithm recognises the relationship between input and result and transfers it to new data.

Typical tasks are classification (Is this email spam?) and regression (What is the expected revenue?). Practical example: banks assess the creditworthiness of applicants.

Unsupervised learning

The model receives no labels – it finds structures and groups in the data itself. Nobody specifies what the algorithm should look for.

Typical tasks are clustering (dividing customers into segments) and dimensionality reduction (simplifying complex datasets). Online shops use the method to identify customer groups with similar buying behaviour.

Reinforcement learning

An agent learns by trial and error in an environment. It receives a reward for good decisions and a penalty for bad ones.

This method controls robots, optimises game strategies and trains autonomous vehicles. The training of ChatGPT also uses reinforcement learning – there, human feedback rates the quality of the answers.

Distinction

What distinguishes machine learning from artificial intelligence?

Artificial intelligence is the umbrella term for all systems that replicate human intelligence – machine learning is the subfield that learns from data.

Hierarchy: AI, ML and deep learning

Artificial intelligence includes machine learning, and machine learning includes deep learning. Not every AI uses machine learning. An expert system that works through hard-coded rules counts as artificial intelligence – but does not learn on its own.

Recognising patterns instead of following rules

Machine learning differs from classic AI in one point: the system programs itself by recognising patterns in data. Classic AI needs a human who defines every rule manually.

Deep dive

What is the difference between machine learning and deep learning?

Deep learning is a specialisation of machine learning that uses artificial neural networks with many layers – particularly strong with images, language and large amounts of data.

Feature engineering vs. automatic feature recognition

In classic machine learning, developers select the relevant characteristics (features) manually. An image recognition model needs, for example, the hint to pay attention to edges or colours. Deep learning skips this step: the neural network recognises for itself which features are relevant.

More performance, more effort

Deep learning needs considerably more training data and computing power. With small datasets or clear tasks, classic ML methods often deliver equally good results – with less effort.

Context

How are machine learning and large language models connected?

Large language models such as ChatGPT and Claude are based on deep learning – a subfield of machine learning that trains neural networks with billions of parameters.

Transformer architecture as the foundation

The transformer architecture forms the foundation. It processes all words of a text simultaneously instead of one after another and recognises via the self-attention mechanism how words relate to each other.

Context: AI, ML, deep learning and LLMs

Artificial intelligence includes machine learning, machine learning includes deep learning, and LLMs are a specific application of deep learning in the field of generative AI. Anyone who understands machine learning understands the foundation behind ChatGPT, Claude and all other language models.

Practice

Where do companies use machine learning?

Companies use machine learning wherever large amounts of data accumulate and patterns need to be recognised faster than humans can. Concrete AI use cases range from fraud detection to predictive maintenance.

Image and speech recognition

Computer vision identifies faces, objects and traffic signs in images. Speech recognition converts spoken language into text – a task from the field of natural language processing and the basis for assistants such as Siri and Alexa.

In medicine, ML models analyse X-ray images and CT scans and detect anomalies that escape the human eye.

Recommendation systems

Netflix, Amazon and Spotify analyse the behaviour of their users and recommend matching films, products or songs. The algorithms learn from clicks, purchases and ratings what the individual user likes.

Fraud detection and risk assessment

Banks and insurers use ML models to detect unusual transaction patterns in real time. A credit card payment at 3 a.m. in a foreign country? The system raises the alarm immediately.

Forecasts and process optimisation

Demand forecasts help retailers plan inventory. Predictive maintenance detects when a machine will fail – before it actually comes to a standstill.

Limitations

What limits does machine learning have?

Machine learning learns only as well as its data – and reaches its limits where data is missing, distorted or decisions have to be explainable.

Data dependency

Without sufficient high-quality data, even the best algorithm delivers poor results. The principle is called “garbage in, garbage out”: faulty or incomplete training data produces faulty predictions.

Bias

Distorted training data leads to discriminatory predictions. A well-known example: application screenings that systematically disadvantage certain groups because the historical data already contained an imbalance.

Explainability

Complex models – especially in deep learning – work as a “black box”. They deliver a result but do not explain how they got there. In regulated industries such as medicine or finance, however, traceability is mandatory.

Energy consumption

Large models need a lot of computing power. Training and ongoing operation cause high CO2 emissions – a growing problem the larger the models become.

Outlook

How is machine learning developing?

Three trends shape the next phase: ML is becoming more automated, runs on end devices and merges with other AI disciplines.

AutoML

AutoML automates model selection, feature engineering and hyperparameter tuning. Companies without their own data science team can thus develop their own ML models – without writing a single line of code themselves.

Edge ML

Models run directly on smartphones, sensors and IoT devices instead of in the cloud. That reduces delays and protects sensitive data because it never leaves the device.

Multimodal models

ML systems process text, image, audio and video simultaneously in a single request. The boundary between speech, image and audio recognition is dissolving.

FAQ

Frequently asked questions about machine learning