Glossary of Terms Related to AI
Glossary of Terms Related to AI
Algorithm, learning algorithm
In Computing Science, an algorithm is a precise calculation rule for solving a task. A learning algorithm is an algorithm that receives sample data (learning data or training data) and calculates a model for the observed data that can be applied to new sample data.
Assistance systems
Digital assistance systems optimise collaboration between humans and computers. They can be found in numerous fields of activity: from document management in the commercial sector, to voice assistants that answer questions or take instructions, through to production and assembly, where they use methods of artificial intelligence support people depending on the context.
Autonomous systems
Autonomous systems are devices and software systems that act and react independently without human control and without pre-programmed processes. They are to be distinguished from automated systems that carry out predefined sequences of actions but cannot change them independently. Autonomous devices must have sensors and software systems that monitor digital data streams in order to react to specific situations. The behaviour is usually determined by machine learning and can be continuously improved.
Basic models or foundation models
Base models are large machine learning models that have been trained on the basis of a large amount of general data. After this pre-training, the models can be fine-tuned for a variety of specific tasks.
A well-known example of foundation models are large language models or large language models (LLMs)which have billions of parameters and can handle complex NLP tasks such as text classification, text generation, language translation, sentiment analysis and question-answer systems. In addition to their use in language models, there are also visual and multimodal foundation models that generate images from text, for example.
Bias
In relation to artificial intelligence, bias refers to a systematic distortion in the predictions or decisions of an AI model. This can occur if the training data is insufficient or unrepresentative. A model with bias can provide incorrect or unfair results. It is crucial to identify and reduce bias in AI systems to ensure that applications are fair and ethically responsible.
Big Data
A large amount of data that is difficult to process using conventional methods due to its size, complexity and speed. This data can be in different formats - structured, unstructured or semi-structured - and comes from a variety of sources such as social media, sensors and online transactions. Big data is often used in conjunction with AI to extract patterns, trends and insights from the data.
Chatbot
A computer programme developed to conduct human-like conversations with users via text or voice. Machine learning methodsare used to train the bots so that they are able to respond adequately to the requirements of future users. Chatbots are often used in customer service, entertainment or other applications.
ChatGPT
GPT stands for Generative Pretrained Transformer. ChatGPT is a powerful language model designed to interact with users in dialogue form. It generates coherent and contextually relevant responses across multiple dialogue steps, which is particularly useful for answering follow-up questions. AutoGPT and BabyAGI are software systems that serve as AI agents and can automate tasks in natural language at a high level. They use extensive language models such as ChatGPT to break down complex tasks into sub-problems that can then be solved automatically using other tools such as a calculator or a search engine.
Computer Vision
An area of artificial intelligence that enables machines to interpret and analyse visual information.
Data Mining
Data mining is the application of statistical and machine learning methods. machine learning to detect patterns, trends or correlations in existing data sets.
Data protection impact assessment (DPIA)
An assessment is required whenever a processing activity, in particular one involving the use of new or emerging technologies, is likely to result in a high risk to the rights and freedoms of natural persons.
Deep learning (DL) or deep learning
Deep learning is machine learning in artificial neural networks with several to very many layers composed of a large number of artificial neurons. Deep learning is responsible for successes in speech, text, image and video processing.
Delusion
In the context of AI, "delusion" refers to a false or misleading idea generated by an AI model, e.g. when the model makes false assumptions about the world or makes inaccurate predictions, see also 'hallucinating'.
Diffusion models or diffusion models
Diffusion models can generate data that is similar to their training data. As generative AI models they are able to generate images based on a text prompt. This is achieved by adding Gaussian noise to the training images and training the model to denoise the image again. The trained model can then generate an image from random noise that resembles its training images.
Discriminative AI or Discriminative AI
Discriminative AImodels learn to differentiate and classify data. In contrast to generative AI modelswhich generate new data, discriminative models assign input data to categories known to them, e.g. animal images to images of dogs or cats.
DPIA (data protection impact assessment)
The data protection impact assessment (DPIA) is a procedure that identifies and evaluates risks associated with the processing of personal data. For digital services and AI applications in particular, it serves to recognise potential data protection risks at an early stage and implement suitable protective measures.
Explainable AI or Explainable AI
Black box models, such as deep neural artificial neural networksare incomprehensible to humans. Explainable AI looks for ways to make the hidden logic or the individual outputs more comprehensible or explainable.
Ethics in AI
An area that addresses the ethical issues and challenges that can arise from the use of AI technologies, such as data protection, privacy, justice and social impact. Consideration of ethical principles is crucial to ensure the responsible use of AI. The research field of AI ethics deals with the development and identification of socially accepted values, principles and techniques as moral guidelines for the responsible development and use of AI systems. The sub-areas of AI ethics include machine ethics, data ethics and the moral behaviour of humans in the design, programming, use and treatment of AI.
Generative AI or Generative AI
Generative AImodels are used to generate new data that has similar statistical properties to a given data set. For example, text, images, audio, video, programme code, 3D models or simulations can be generated that follow the user's instructions.
Large language models (LLMs)
Large language models are Foundation Models or foundation models that have been trained to process natural language with large amounts of text data. The models learn to continue texts by establishing statistical relationships between words, thus building up knowledge about the syntax, semantics and ontology of the language. After this pre-training, the models can be fine-tuned for their specific use, e.g. as a chatbot. Your transformer-architecture enables the efficient processing of large amounts of data and the consideration of remote dependencies in data.
Hallucinating
Generative AI uses extensive data sets to recognise patterns and create new content, be it in image, audio or text form. These patterns serve as the basis for generating new responses that are similar to the previous database. However, despite the plausibility of the generated content, errors can occur that lead to so-called "hallucinations". These inaccuracies result from the quality of the training data and can lead to misleading results. It is important to emphasise that the quality of generative AI strongly depends on the quality of the training data. Incorrect or misleading data can lead to inaccurate or erroneous results. Therefore, a responsible approach to generative AI is essential by carefully checking data sources and critically scrutinising the results to avoid potential misinformation.
HRKI systems (high-risk AI systems)
HRKI systems refer to AI applications whose use can entail a particularly high risk in terms of health, safety or the impairment of fundamental rights. Particularly in the context of educational applications, the special duty of care is emphasised in order to avoid negative effects.
Hybrid AI
Hybrid AI combines data-based machine learning, knowledge representation and logical reasoning. Knowledge and the respective conclusions are incorporated directly into the learning process, for example to emulate the human ability to correctly understand meanings from the context and to make the AI system more robust overall.
AI systems with a general purpose (GPAI (General Purpose AI))
General-purpose AI systems refers to general-purpose AI models and systems that can be used flexibly in a wide range of applications - from text generation to image and video processing. The regulatory framework examines the extent to which these systems pose potential systemic risks.
AI Regulation (AI Act)
The AI Regulation - also known as the AI Act - is an EU regulation that governs the safe and responsible use of artificial intelligence. The aim is to promote innovation while ensuring high standards of protection for health, safety and fundamental rights.
Cognitive machines, cognitive systems
Cognitive machines or systems are alternative terms for artificial intelligent systems or artificial intelligence. artificial intelligence. They are characterised by learning and reasoning capabilities as well as language processing, image processing and interaction with the user.
Conversational AI
AI systems that are designed to understand and generate natural language in order to have conversations with users. Conversational AI includes chatbots, voice assistants and other applications that use natural language processing technologies.
Artificial intelligence (AI) or artificial intelligence (AI)
Artificial intelligence is a branch of Computing Science that deals with the automation of intelligent behaviour. There is no fixed definition of what "intelligent" means, nor which technology is used. One of the foundations of modern artificial intelligence is machine learning. Other important methods are logical reasoning based on symbolic knowledge, knowledge representation or planning procedures. Experts distinguish between strong AI and weak AI.
Artificial neural networks (KNN)
Artificial neural networks are models of machine learning. machine learningmodelled on the natural neural networks of the brain. They consist of many layers of nodes realised in software, which are referred to as artificial neurons. With the help of examples, a learning algorithm changes the weights, numerical values at the connections between the nodes, until the results are good enough for the task. The number of nodes, layers and the links between them have a significant effect on the model's ability to solve the problem.
Natural language processing (NLP)
A technology that enables computers to understand and respond to human speech.
Machine learning (ML)
Machine learning aims to generate knowledge from empirical values by learning algorithms develop a complex model from examples. The model can then be applied to new, potentially unknown data of the same type. This means that machine learning does not require manual knowledge input or explicit programming of a solution path.
Machine language processing or natural language processing (NLP)
Machine language processing comprises techniques for recognising, interpreting and generating natural language in spoken and written form. This includes the textualisation of spoken language, mood recognition, information extraction from texts, machine translation and conducting conversations.
Model
A model is an abstraction of reality. In machine learning, a learning algorithm generates a model that generalises the fed-in data. The model can then be applied to new data.
Multimodal AI
While unimodal AI systems can only process or generate one type of data, multimodal AI can handle different types of data such as text, images and audio. Multimodal models are therefore more flexible as they are trained with different data types.
Neural network
A neural network is a computer-based model inspired by the human brain. It consists of numerous artificial neurons that are arranged in layers and connected to each other. Each neuron receives information, processes it and forwards the result to other neurons until a specific goal is achieved. Neural networks are used in various artificial intelligence techniques such as deep learning to perform complex tasks such as pattern recognition, classification and prediction.
Open source models
AI models or algorithms whose source code is publicly accessible and can be freely used, modified and improved by the community. Open source models promote collaboration and knowledge sharing in AI research and development.
Personal data
Any information relating to an identified or identifiable living individual, including various data points which, when combined, may lead to the identification of a specific individual.
Predictive analytics
Predictive analytics is an area of artificial intelligence that focuses on predicting future events or trends by using existing data and specialised techniques. By analysing past data, it identifies patterns and then makes predictions about upcoming developments. This approach is used in various areas such as marketing, finance, healthcare and logistics and enables intelligent decisions to be made and risks to be minimised.
Prompt
In artificial intelligence, a "prompt" refers to a text or instruction given to a language model to fulfil a specific task or generate an answer. This can be a question, a description or a sentence. The model is then instructed to complete the missing part of the text. The quality and clarity of a prompt often have a direct influence on how precise and relevant the model's generated response is.
Reinforcement learning
With reinforcement learning, the learning algorithm receives occasional feedback for interactions with the environment and learns to better assess the chances of success of individual actions in different situations. Reinforcement learning is often used for autonomous systems and games.
Robots
Robots are machines or devices that are designed to take over certain physical and communicative tasks from humans. Typical examples are service and industrial robots. The autonomy of robotic systems increases to the extent that they can perform tasks independently through machine learning to solve complex tasks. One example of this is fully autonomous vehicles.
Weak AI
Weak AI uses AI methods to solve narrowly defined tasks. While it can already outperform human capabilities in individual areas, such as image analysis, weak AI is nowhere near the same level for broader tasks in a wider context or for tasks that require knowledge of the world. All current AI solutions are examples of weak AI.
Security and safety-by-design
Safety-by-design and security-by-design are principles that aim to ensure that security requirements are taken into account from the outset when developing both software and hardware for (AI) systems. The aim is to avoid potential vulnerabilities and opportunities for attack. While security-by-design aims to prevent criminal attacks, safety-by-design focuses on avoiding accidents and other safety-related risks.
Strong AI or Artificial General Intelligence
Strong AI stands for the vision of using AI technologies to replicate human intelligence in its entirety and outside of individual, narrowly defined fields of action. So far, strong AI has only been found in science fiction. Ever since artificial intelligence emerged in the 1950s, there have been predictions that strong AI would be realisable within a few decades.
Stochastic parrot
The term "stochastic parrot" describes a criticism of large language models and refers to the way in which they generate texts: They imitate human language by relying on probabilities and recognising patterns in huge data sets without having any real understanding of the content. The analogy shows that such models, like a parrot, can reproduce meaningful and convincing sentences, but have no actual meaning or contextual understanding. The term thus emphasises the limitations of purely statistical models in terms of semantic depth and critical thinking.
Superalignment
A concept in AI ethics that aims to design AI systems in such a way that they not only achieve their intended goals, but also act in accordance with people's values, goals and preferences. Superalignment calls for a comprehensive alignment of AI systems with human interests and values.
Token
A digital unit that acts as a representation of an asset, a unit of value or a right to access a resource. In relation to AI, tokens could be used to gain access to AI services, data or applications, or as a reward for participating in AI-powered platforms - a token can correspond to a word syllable, for example.
Transformer
A Transformer is a deep learning-architecture that uses anattention mechanism to map relationships between words. Due to the efficient processing of large amounts of data and the consideration of remote dependencies in data, transformer models are used in machine language processing for understanding, translating or generating texts, but also in image processing. Transformers are best known for their use in large language models.
Trolley problem
The classic trolley problem is particularly relevant in the field of autonomous driving. It is a philosophical thought experiment that depicts a dilemma situation in which both courses of action lead to undesirable outcomes. In this scenario, a driverless tram hurtles inexorably towards five people chained to the track. By actively changing a switch, the tram could be diverted to another track, but with one person chained to it. The decision as to whether the switch should be moved presents those involved with a moral dilemma that raises ethical questions about weighing up actions and their consequences. The aim of this thought experiment is to encourage reflection on complex ethical decisions and the morality behind such situations. This thought experiment is repeatedly used as a moral compass for AI.
Turing test
The Turing test was designed by British mathematician Alan Turing to assess the intelligence of artificial systems. If a person communicating with an artificial system and a human dialogue partner at the same time is ultimately unable to determine which dialogue partner is the human, the system is considered intelligent. Nowadays, such systems are referred to as chatbots.
Supervised learning or supervised learning
In supervised learning, the training data consists of examples with input and output. The model should learn a function in order to predict new examples well. To determine the quality of the model, it is trained with only part of the available data and the finished model is tested with the remaining data.
Distributed AI or distributed AI
With machine learning in the cloud, the model is model only in the cloud. To train and apply it, the end devices have to send all the raw data to the server. With distributed AI, the models remain in the end devices. Instead of the raw data, the models are uploaded to the cloud, combined there and distributed again. In this way, each end device benefits from the training on all other end devices. The data protection-friendly concept of edge computing goes hand in hand with savings in computing time, communication effort and costs, as well as increased security against cyberattacks.
Trustworthy AI or trustworthy AI
Only trustworthy AI applications guarantee IT security, control, legal certainty, accountability and transparency. For this reason, guidelines for the ethical design of artificial intelligence are being developed within the company and at a social and political level. artificial intelligence are being developed. These focus, for example, on the dimensions of ethics and law, fairness, autonomy and control, transparency, reliability, security and privacy.
Knowledge representation
Different methods of knowledge representation are used to formally represent knowledge, e.g. ontologies, classes or semantic networks or rule systems. The expert systems of the 1980s consisted of such knowledge bases. Today, rule systems are often used for programming chatbots.
Zero-shot learning
A method of machine learning in which a model is able to master tasks without having been explicitly trained for these tasks beforehand. It uses existing knowledge to solve new, unknown tasks.
Certification
Current efforts to develop a test catalogue for AI applications aim to enable the certification of AI applications. The standards set in this way are intended to make the quality of AI applications assessable in a differentiated manner, contribute to transparency in the market and promote acceptance in the application.