Glossary

Machine learning is a programming method that allows computer systems to learn from the data provided. We call a computer system learning if it performs its task increasingly well as data - called learning data - comes in.

With supervised machine learning, the teaching data is prepared by a human and contains the expected results of the task. For example, images are prepared for an image recognition task along with information about what is in them - the more such input, the better the system performs its task.

Unsupervised machine learning is based on the fact that the system learns from large data sets devoid of the expected results of a specific task. The system learns independently to capture the relationships present in the data. For example, an algorithm for grouping similar objects works effectively without prompts.

In reinforcement machine learning, the system learns by interacting directly with the environment in which it operates, actual or simulated, receiving feedback in terms of rewards for desired responses and penalties for undesired responses. For instance, an autonomous driving system can be taught (in a laboratory setting) with rewards for reaching a destination and penalties for causing incidents.

Artificial intelligence (AI) is a computer science discipline that deals with the development of computer systems simulating human intelligence. This simulation may involve:

• the learning process implemented by machine learning,
• input data processing, given in a form intelligible to the human senses, mainly vision and hearing, e.g. texts, images, video recordings, sounds or human speech,
• generate output data given in a form intelligible to humans.

Generative artificial intelligence (GSI, or GenAI) is a type of artificial intelligence that focuses on generating output given in a form understandable to humans. Based on patterns captured in the learning data, systems of this type can create new, original entities. Examples of systems include ChatGPT, generating natural language text, Midjourney generating images resembling the works of a human painter, Copilot creating programming code, or Sora generating videos with user-set content.

A neural network is a computational method in which a set of 'neurons' (computational units inspired by nerve cell functioning), networked together, transform and analyse data similarly to the human brain. The neurons are arranged in layers: input, output and hidden layers in between. Each successive layer receives data resulting from the data processing in the previous layer, the input layer being the input data (e.g. information about the image to be recognised) and the output layer representing the output data of the system (e.g. information about what object has been recognised in the image).

A deep neural network is a neural network that contains a large number (usually several dozen to several thousand) of hidden layers.

Deep learning is a machine learning type implemented using deep neural networks.

The number of parameters defines the number of interconnections between neurons in a neural network and determines its size.

A Large Language Model (LLM) is a deep neural network trained to generate a natural continuation of an utterance learned on enormous text datasets. Through the use of supervised machine learning and machine learning by reinforcement, large language models have been adapted to tasks such as answering questions, conducting dialogue, producing prose or generating programming code.

Examples of large language models include the American ChatGPT (in the latest version 4.0, available free of charge, with restrictions), Claude (available free of charge, with restrictions), Gemini (available free of charge provided you log in to Google's service), Llama (with free access and teachable on your data) and the European-developed Mistral (for smaller versions with a licence similar to Llama). The parameter numbers of these models range from a few to several hundred billion.

Hallucination is the unwanted action of a large language model in generating information that is false, nonsensical or unrelated to the query. For example, a grand language model asked to list a bibliography for a given topic may generate titles of non-existent, though realistic-sounding, books or articles.

Bias is the undesired action of a generative AI system to generate biased, unbiased or discriminatory content, e.g. politically, racially or ideologically motivated content. Bias is usually caused by inadequately prepared learning data that does not include a complete diversity of examples of all groups or phenomena. An example would be a film recommendation system - if it relies on data mainly from young adults, it may offer recommendations that are less suitable for older adults.

A token is the smallest unit of language interpreted by a large language model. For example, the word 'intelligence' is interpreted by the ChatGPT model as consisting of two tokens: 'inteligen' and 'cja'.

A prompt is a user instruction directed to the grand language model to obtain the desired information. The prompt is usually given as a natural language query or request, e.g. 'Explain what the term artificial intelligence means'. The length of the prompt generally ranges from a few words to several sentences. The quality (compliance with expectations) received depends on the proper preparation of the prompt.

A large language model can respond to a prompt in a non-deterministic manner - it can generate a different response to the same user utterance at various times.

The context window is the size of the text, determined by the number of tokens, that is taken into account by the grand language model in generating the utterance. By creating a response to a given prompt, the grand language model can refer to the conversation history between the user and the system, the length of which does not extend beyond the context window.

An AI detector is a system designed to detect content generated by artificial intelligence systems, specifically by grand language models. The performance of AI detectors can vary significantly depending on the language or domain of the text under investigation and on a version of the large language model used to generate the text under investigation. The results reported by AI detectors should be approached with great caution, as there is no guarantee that the information returned by AI detectors is correct.