AI Glossary

  • Algorithm: A set of formulas and/or instructions given to a computer for it to complete a task.

    Artificial Intelligence (AI): The field of computing that aims to create machines or software capable of intelligent behavior. AI systems can perform tasks that typically require human intelligence, such as visual perception, speech recognition, decision-making, and language translation.

    AI Architecture: The architecture is like the blueprint of an AI system. The design of an AI system could vary in terms of the type of problem-solving methods (i.e., class of algorithms) used, the way in which the system handles information (i.e., data management techniques such as data preprocessing, data cleaning, data transformation, and data storage methods), and/or the computational resources (i.e., bandwidth, memory, storage, hardware) required. A common type of AI architecture is the transformer class of models, which is used in the GPT (Generative Pretrained Transformer) models. Each architecture has its own strengths and limitations for particular uses.

    AI Hallucination: A hallucination occurs when an AI system produces output that may sound plausible and coherent but is factually incorrect or based on faulty assumptions. While AI doesn’t have agency and should not be thought to act in a human-like way, the term is intended to highlight the phenomenon whereby the output of an AI system may be misleading and could contribute to misinformation. Hallucinations can be caused by an AI system being trained on inaccurate data, incomplete data, and/or data that only explains a narrow set of phenomena. It is important to be aware of the limitations of AI, and to know that output should be checked for factual accuracy.

    AI Literacy: The ability to use and interact with AI systems effectively, efficiently, and responsibly.

    Bias in AI: The systematic prevalence of untrue and/or harmful information that leads to unfair, inaccurate, or discriminatory outcomes in the decision-making process or results of an AI system. It can manifest directly or indirectly, stemming from the data used to train the system or the design of the algorithm itself.

    Data: Text or numeric information that can be stored, processed, or analyzed by a computer.

    Data Privacy: The aspect of handling, processing, and storing personal and sensitive data with confidentiality and security. In AI, this involves ensuring that data used for training and operating AI systems is protected from unauthorized access and misuse.

    Deep Learning: Deep learning is an advanced type of machine learning. It studies numerous examples (like images or speech) and translates them into layers of numbers and statistical models, often referred to as ‘neural networks’, to capture complex patterns in data. These neural networks mimic the way our brain works. Deep learning is particularly effective for tasks like image and speech recognition.

    Ethics in AI: The moral principles and guidelines that govern the development and use of AI technologies. This includes considerations of fairness, transparency, accountability, and the impact of AI on society and individuals.

    Generative AI: A type of AI that specializes in creating new content, such as text, images, and audio, by learning from existing data. It can generate realistic and coherent outputs that mimic original data sources.

    Graphics Processing Unit (GPU): A GPU is a chip/semiconductor used by a computer to increase the speed with which a computer can run calculations, such as AI models that require deep learning. The GPU can use hundreds or even thousands of smaller “cores”, which are analogous to individual employees completing tasks within a company, to run processes in parallel. While the central processing unit (CPU) is the main hardware that runs a computer, a GPU can take many instructions (including those that have been split into smaller parts) and run them all at the same time instead of sequentially. GPUs are relevant in many contexts outside of AI (e.g., graphics acceleration for video games) but have been particularly advantageous for increasing how quickly AI models can be trained and implemented because of the parallel processing, dedicated high-speed memory (known as VRAM), and the ability to run large volumes of small calculations (such as those needed in matrix multiplication for AI model calculations).

    Large Language Model (LLM): Complex statistical models that use billions of variables to capture the characteristics and content of language across billions of sentences. LLMs can be used to study how the English language is written across different contexts and use that understanding to write new text (e.g., answering questions, creating summaries or lists).

    Machine Learning (ML): The ability of a machine to predict outcomes without someone giving it exact instructions. ML involves complex statistical models to make predictions (i.e., learning the connections between variables).

    Model: In AI, a model refers to the specific trained algorithm capable of performing tasks such as prediction, classification, or content generation. Models are trained using datasets and refined until they achieve the desired level of accuracy.

    Multimodal AI Model: An AI system that can process multiple modalities, or types, of data (e.g., text, images, audio video, environmental) and/or create multiple types of output.

    Natural Language Processing (NLP): A branch of AI that focuses on the interaction between computers and humans through natural language. NLP involves the interpretation, understanding, and generation of human language by AI systems.

    Neural Networks: Computational models that are inspired by the human brain's structure and function. These networks are composed of interconnected nodes (similar to neurons) and are used to model complex patterns in data for tasks such as classification and prediction.

    Objective-Driven AI: An AI architecture that is modular and designed to meet a set of pre-defined objectives created by people, which include guardrails to ensure controllability and safety. It constructs internal models of how the world works (i.e., “world models”) by learning from experiences and uses them to predict how to learn and respond in new or uncertain situations. While this architecture can be used to process text data, it is more likely to use videos and/or sensors to study the environment around it.

    Prompt: A question or task provided to an AI tool that is the basis for the action it should perform. A prompt generally contains detailed information related to what the computer should do and any constraints under which it should operate. 

    Small Language Model (SLM): An AI system that addresses similar tasks as a Large Language Model but uses far fewer parameters to do so. One advantage of SLMs is that they can run locally on devices (i.e., connection to a cloud environment may not be needed) and may be better suited for even faster response times and better protecting data privacy. It may be expected that SLMs will be used alongside LLMs in various applications, where SLMs may be trained for specific use cases and used as needed by a larger AI workflow.

    State Space Model (SSM): An AI architecture that uses a different class of statistical modeling than LLMs. State space models analyze sequences of data (e.g., words within a sentence) by using the current “state” of the data (i.e., how it currently exists) and how it changes over time. These models are more frequently applied to multimodal data that span multiple time points.

    Training Data: The dataset used to train AI models. This data provides the examples and experiences that AI systems use to learn how to perform tasks, make predictions, or generate content. The quality and diversity of training data significantly influence the performance of AI models.