A compilation of frequently used terms relevant to systems of Artificial Intelligence.
Printed terms are cross-references to other terms that are also explained in the AI glossary.
Further information on AI can be found in contributions on Dr. GDPR:
Dictionary of important AI terms
| Term | Explanation |
|---|---|
| Artificial Intelligence | Also abbreviated as AI. AI is the development of computers and software that possess human-like abilities such as learning, problem-solving, language understanding and production, sensory perception, action control, and autonomous decision-making. |
| Model | An AI model is an electronic brain. Just like in a biological brain, it consists of a neural network (see ANN). |
| Large Language Model | Abbreviation for Large Language Model. A model that has been trained on a large number of text documents in order to be able to represent the grammar of one or more languages very well. |
| Foundation Model | A well-trained foundational model that can be fine-tuned for specific tasks or domains. Such models are often provided under a user-friendly license. |
| ANN | Abbreviation for Artificial Neural Network. Electronic representation of the biological model. Basis for AI models (see Model). |
| Translator | A crucial mathematical approach that has made high-performance AI applications like ChatGPT possible, and is responsible for the current success of AI. Exists since 2017. The precursor was developed by Jürgen Schmidhuber in the early 1990s. Consists of an Encoder-Decoder structure, where input is encoded into features that are then decoded into output. See also Embeddings. |
| Model size | The size of a AI model (see model) is given in shorthand form by the number of parameters (see parameters). Small models have only several hundred million parameters. A 200M model has 200 million parameters. Larger models have 13B or 60B parameters. The "B" stands for billion (English: "billion"). |
| Parameter | As parameters, weights are called in an artificial neural network (see ANN). A weight defines the strength of the connection between two neurons. |
| ChatGPT | ChatBot based on GPT architecture (see GPT). Also has a programmable interface (API). Should have approximately 160B parameters (cf. model size). Problematic with regards to data protection, business secrets and confidential documents. |
| GPT | Abkürzung für Generative Pretrained Transformer. Siehe Generativ, Vortrainiert und Transformer. |
| Generative | Ability to create or produce something new. Part of the abbreviation GPT, which is also embedded in ChatGPT. |
| Well-trained | A pretrained model is a trained model. It has been trained. Instead of saying "trained", however, we say "pretrained". The background is that an already trained model can be further trained, which is called fine-tuning. |
| Fine-Tuning | Pre-trained models (see pre-trained) can be further "trained" for specific tasks (see downstream). This further training is called Fine-Tuning. Fine-Tuning has the great advantage that much less input data and resources (time, computational power, main memory) are required than to build a pre-trained AI model. |
| Layers | Refers to artificial neural networks (see ANN): The layers of a neural network are the input layer, the output layer, and the hidden layers between these two layers. |
| Input layer | First layer of an artificial neural network (see ANN). Input data (sensor data or user input) reaches the ANN through the input layer. |
| Output layer | Last layer of an artificial neural network (see ANN). The output layer announces the result of a machine learning algorithm. |
| Hidden Layers | Layers of a ANN, situated between input layer and output layer. It is here that magic happens, also referred to as intelligence. See human intelligence and artificial intelligence. |
| Hidden layers | See Hidden Layers. |
| Upstream Task | Task for which an AI model has been pre-trained. See also Downstream Task. |
| Downstream Task | Task for which a AI-model is prepared with the help of Fine-Tuning and which differs from the Upstream Task for which the AI-model was originally trained. |
| PyTorch | Most popular AI framework, based on the programming language Python. |
| Tensorflow | Besides being a well-known AI framework, PyTorch is from Google. It's considered more complex than PyTorch. |
| Array | Mathematical concept for data representation. See also Vector. |
| Delegation | Offloading parts of a machine learning model (see Model) onto a CPU (instead of GPU) or hard drive. Offloading resolves memory issues (see VRAM) with large ML models, but results in significantly slower computations. |
| CUDA | Abbreviation for Compute Unified Device Architecture. Software architecture of Nvidia, to use graphics cards (see GPU) for calculations. |
| Nvidia Corporation | Without wanting to advertise: World-leading provider for graphics cards that are particularly well-suited for AI calculations. See CUDA. There is no other manufacturer I know of whose graphics cards are supported as well in numerous AI frameworks (see PyTorch). |
| Python | Most popular programming language for AI applications. |
| Graphics Processing Unit | Abbreviation for Graphics Processing Unit. A GPU of a high-performance graphics card like from Nvidia contains thousands of cores. A core can solve a calculation task. Several cores can work simultaneously. For AI applications, billions of calculations can be executed on a GPU (graphics card) several times faster than on a CPU (normal processor). The graphics card is misused as a calculator in AI applications. It does not output images or text. |
| CPU | Abbreviation for Central Processing Unit. The processor of a computer. It typically has 8 to 24 cores. See also GPU. |
| Video Random Access Memory | Video RAM. In contrast to RAM, here the memory of a graphics card (see GPU) is meant. Decisive for the use of AI models (see Model). |
| AI Server | This is the name given to servers that contain a powerful graphics card (see CUDA, GPU and VRAM). |
| Stable Diffusion | One of the most well-known and popular AI models for image generation. |
| LAION | Abbreviation for Large Scale Artificial Intelligence Open Network. A registered association in Germany. Has created a dataset of nearly 6 billion images along with image descriptions. Serves as a basis for Stable Diffusion approaches. |
| Image creation | Most common application: A picture is generated by AI based on a prompt, which reflects the prompt. The best-known procedure is Stable Diffusion. It's also possible to combine multiple images or modify input images. |
| Vector | A vector is a representation of an input (data set), such as a text, image or audio file. See multimodal. A vector consists of number sequences and has a dimension that defines the number and arrangement of numbers. |
| Embedding | Representation of data of any kind (text, image, audio, motion sensor, temperature sensor, etc.) in the form of a vector. |
| Multimodal | Thanks to the Transformer approach, any type of data can be treated equally. Compare the human brain. All sensory data (eyes, ears, nose, skin stimuli…) are "naturally" processed in our brain's neural network in the same way. |
| Human Intelligence | Works like Artificial Intelligence, that's my strong thesis. |
| Spirit | Illusion, which is produced by the human brain, is my conviction. |
| Artificial General Intelligence | Abbreviation for Artificial General Intelligence. Analogous to humans: general intelligence that can solve all possible tasks. |
| Prompt | Input field for text, to give instructions to an AI or ask a question. |
| Token | A data unit, such as a word in a text. Common unit for calculating the cost of using AI systems like ChatGPT or configuring the maximum output length of text. |
| Long Short-Term Memory | Abbreviation for Long Short Term Memory. In German: Long-term short-time memory. Approach conceived by German computer scientist Jürgen Schmidhuber. LSTM is a type of episodic memory. |
| Statistics | In AI-models equally represented or less than in humans (see human intelligence) |
| Artificial Intelligence | Refers to the fact that a ANN has many hidden layers, so it's deep. |
| Self-sufficient system | Friendly data system that is operational on a local infrastructure and without internet connection. |
| Data-friendly | System, that does not cause homemade problems with data protection, confidential data or business secrets. The opposite are systems that come from providers in the USA or hold data on servers in the USA (cf. Privacy Shield – Schrems II). |
| Eliza | Psychiatrist program that was conceived as early as 1966 by Joseph Weizenbaum. Pretended to be intelligent. Instead, it conveyed the feeling of intelligence to the human conversation partner by using clichés and replaying back the human's prompt inputs. |
I gladly refer also to the Data Protection Dictionary German-English, which contains the most important terms for the DSGVO and digital data protection.
Key messages
Artificial intelligence (AI) uses computer systems to mimic human abilities like learning and problem-solving by using complex algorithms and large datasets.
Fine-tuning AI models requires less data and resources than training them from scratch.
This text explains common terms related to artificial intelligence (AI), focusing on how AI models process information and generate outputs like images based on user prompts.
My name is Klaus Meffert. I have a doctorate in computer science and have been working professionally and practically with information technology for over 30 years. I also work as an expert in IT & data protection. I achieve my results by looking at technology and law. This seems absolutely essential to me when it comes to digital data protection. My company, IT Logic GmbH, also offers consulting and development of optimized and secure AI solutions.
