Artificial intelligence and machine learning

Artificial intelligence (AI) is the field of creating programs that can do tasks which need human intelligence, such as recognizing images, understanding speech, writing text or making decisions. In an ordinary program the programmer writes all the rules. In machine learning the program finds the rules itself by analyzing many examples. Such programs are used today in translators, voice assistants, recommendation systems and medical diagnosis.

Machine learning has several stages. First, examples are collected: for instance, thousands of emails labeled “spam” or “not spam”. Some of them are set aside to train the model and the rest to test its quality; a model must be tested on examples it has not seen, otherwise it may have simply memorized the answers. Training on examples with known answers is called supervised learning, while finding groups in data without answers is called unsupervised learning.

Most modern models are based on artificial neural networks: they consist of many simple computing units, and during training the “weights” of the connections between them are adjusted. Large language models are trained on huge amounts of text and produce text by predicting the next word. They can sound convincing yet give wrong answers (“hallucinations”), so important information should be checked in reliable sources.

Ethical questions matter when using AI. A model learns only from the data it was trained on, so if the data is not balanced (for example, some groups are under-represented), the model’s decisions will be biased too. Protection of personal data, copyright and responsibility for decisions all need to be considered. AI is a powerful tool, and being able to study it and use it critically is useful in modern professions.

Key takeaways

• In machine learning the program finds the rules itself from many examples.
• Data is split into training and test parts: a model must be tested on unseen examples.
• In supervised learning the answers are known; in unsupervised learning they are not.
• Neural networks and large language models can sound convincing yet be wrong.
• A model repeats the bias in its data; personal data and responsibility are important issues.