| Active learning |
In the context of artificial intelligence,
1) a reinforcement learning paradigm where the agent explores given states and updates its exploration policy when it receives a reward or avoids a penalty.
2) a form of model-assisted learning wherein the agent learns to infer new data labels from an initial set of labeled data.
3) a specific form of machine learning in which an agent can interactively query a user or another data source to label new data.
This technique is commonly used when there is some cost to acquiring new data, and when one can pick data points to explore which provide the most information. |
Richard S. Sutton and Andrew G. Barto, Reinforcement Learning: An Introduction, 2nd ed. (Cambridge, MA: MIT Press, 2018), http://incompleteideas.net/book/the-book-2nd.html.
David Cohn, "Active Learning," in Encyclopedia of Machine Learning, eds. Claude Sammut and Geoffrey I. Webb (Boston, MA: Springer, 2011), 10-14, https://doi.org/10.1007/978-0-387-30164-8_6.
Chief Digital and AI Office | Chief Technology Office |
Learning Methods and Techniques |
| Adversarial example |
A data instance that is purposefully perturbed to induce misclassification by a deployed model. |
Apostol Vassilev, Alina Oprea, Alie Fordyce, and Hyrum Anderson, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations, NIST Trustworthy and Responsible AI, NIST AI 100-2e2023 (Gaithersburg, MD: National Institute of Standards and Technology, January 2024), https://doi.org/10.6028/NIST.AI.100-2e2023. |
Security |
| Adversarial success |
A state of model breakdown that includes leakage of privileged data, such as a system prompt, context, or training data, or a violation of inserted guardrails due to an adversary's successful attack on a generative artificial intelligence model. |
Apostol Vassilev, Alina Oprea, Alie Fordyce, and Hyrum Anderson, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations, NIST Trustworthy and Responsible AI, NIST AI 100-2e2023 (Gaithersburg, MD: National Institute of Standards and Technology, January 2024), https://doi.org/10.6028/NIST.AI.100-2e2023. |
Security |
| Agent |
In the context of artificial intelligence, a program acting on behalf of a person or organization.
In the context of reinforcement learning, a name for the model in a reinforcement learning process.
In the context of large language models, any program that relies on a large language model to perform complex tasks that require task planning and the use of external tools. |
"Glossary," Computer Security Resource Center, National Institute of Standards and Technology, updated October 17, 2023, https://csrc.nist.gov/glossary/term/agent.
John Hwang, "Learn How to Build and Deploy Tool-Using LLM Agents Using AWS SageMaker JumpStart Foundation Models," AWS Machine Learning Blog, September 15, 2023, https://aws.amazon.com/blogs/.
machine-learning/learn-how-to-build-and-deploy-tool-using-llm-agents-using-aws-sagemaker-jumpstart-foundation-models. |
Applications and Use Cases |
| Alignment |
In the context of generative artificial intelligence, a measure of whether artificial intelligence models represent human preferences or values in their behaviors and outputs. |
Sohl Dickstein, "BIG-Bench Keywords," GitHub, updated June 4, 2022, https://github.com/google/BIG-bench/blob/main/keywords.md. |
AI Ethics and Governance |
| Alignment problem |
In the context of generative artificial intelligence, a situation in which artificial intelligence responses do not correspond to human moral or aesthetic preferences, possibly because the artificial intelligence optimizes for tractable answers to complex, interdependent social problems. |
Brian Christiansen, The Alignment Problem: Machine Learning and Human Values (New York, NY: W. W. Norton & Company, 2020). |
AI Ethics and Governance |
| Annotation |
The creation of labels for data used to train artificial intelligence applications. Also called "data annotation," "labeling," and "tagging." |
"What Is Data Annotation," Appen (blog), July 10, 2020, https://appen.com/blog/data-annotation/. |
Data Handling and Processing |
| Application-specific integrated circuit (ASIC) |
A specialized computer processing chip that improves speed and efficiency for specific tasks, such as neural network calculations. |
"Encyclopedia," PCMag, updated monthly, https://www.pcmag.com/encyclopedia/term/asic. |
Hardware and Computational Resources |
| Artificial general intelligence (AGI) |
Human-like intelligence that can be applied widely. It contrasts with "artificial narrow intelligence," which can be applied only to one particular problem or task. Also called strong artificial intelligence, as opposed to weak artificial intelligence. |
Mark Coeckelbergh, AI Ethics (Cambridge, MA: MIT Press, 2020), https://mitpress.mit.edu/9780262538190/ai-ethics/. |
AI Ethics and Governance |
| Artificial intelligence (AI) |
In the context of computer science, the term is continuously evolving. A recent definition is "a machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations, or decisions influencing real or virtual environments. Artificial intelligence systems use machine- and human-based inputs to perceive real and virtual environments; abstract such perceptions into models through analysis in an automated manner; and use model inference to formulate options for information or action." |
15 U.S.C. 9401(3).
Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence, Section 3(b), October 20, 2023, https://www.whitehouse.gov/briefing-room/presidential-actions/2023/10/30/executive-order-on-the-safe-secure-and-trustworthy-development-and-use-of-artificial-intelligence/. |
AI Ethics and Governance |
| Artificial intelligence red teaming |
In the context of artificial intelligence, this evolving term describes any of a number of specific testing exercises for finding vulnerabilities or bugs in an artificial intelligence system. A red team may adopt any of the adversarial methods known or suspected to cause fault in an artificial intelligence system. |
Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence, Section 3(d), October 20, 2023, https://www.whitehouse.gov/briefing-room/presidential-actions/2023/10/30/executive-order-on-the-safe-secure-and-trustworthy-development-and-use-of-artificial-intelligence/.
Apostol Vassilev, Alina Oprea, Alie Fordyce, and Hyrum Anderson, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations, NIST Trustworthy and Responsible AI, NIST AI 100-2e2023 (Gaithersburg, MD: National Institute of Standards and Technology, January 2024), https://doi.org/10.6028/NIST.AI.100-2e2023. |
Security |
| Artificial intelligence system |
In the context of computer science, the term is continuously evolving. A recent definition is "any data system, software, hardware, application, tool, or utility that operates in whole or in part using artificial intelligence." |
Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence, Section 3(e), October 20, 2023, https://www.whitehouse.gov/briefing-room/presidential-actions/2023/10/30/executive-order-on-the-safe-secure-and-trustworthy-development-and-use-of-artificial-intelligence/. |
AI Ethics and Governance |
| Artificial narrow intelligence (ANI) |
A level of intelligence that exceeds human capability and is specific to a particular problem or narrowly specified task. Compare "artificial general intelligence." |
Eban Escott, "What Are the 3 Types of AI? A Guide to Narrow, General, and Super Artificial Intelligence," October 24, 2017, https://codebots.com/artificial-intelligence/the-3-types-of-ai-is-the-third-even-possible. |
AI Ethics and Governance |
| Attention |
In the context of artificial intelligence, and transformer models more specifically, a technique which efficiently relates different positions within an input sequence, in order to produce information with context |
Sebastian Raschka, "Understanding and Coding the Self-Attention Mechanism of Large Language Models from Scratch," AI Magazine, updated February 9, 2023, https://sebastianraschka.com/blog/2023/self-attention-from-scratch.html. |
Neural Network Components and Functions |
| Attention head |
In the context of artificial intelligence and transformer models more specifically, the ability of a model to take a single combination of key, query, and value matrices that define an attention relation. Multiple attention heads are typically used in a single transformer model. Intuitively, multiple attention heads allow the model to attend to parts of a sequence of terms or tokens in different ways. For example, one attention head may attend to sequences of data points with respect to their longer-term dependencies and another to shorter-term dependencies. |
Ashish Vaswani et al., "Attention Is All You Need," Advances in Neural Information Processing Systems 30 (NIPS 2017), Figure 2, https://doi.org/10.48550/arXiv.1706.03762 |
Neural Network Components and Functions |
| Autoencoder |
"A type of artificial neural network used to learn efficient codings of unlabeled data, typically for the purpose of dimensionality reduction. An autoencoder employs unsupervised learning to learn a representation (encoding) for a set of data, typically for the purpose of reducing the dimensionality of the data. The network is trained to compress the input into a lower-dimensional code and then reconstruct the output from this representation to match the original input as closely as possible, hence the name autoencoder." |
"Machine Learning Glossary," DeepAI, https://deepai.org/machine-learning-glossary-and-terms/autoencoder. |
AI Model Types and Architectures |
| Backpropagation |
A set of algorithms used to train feedforward neural networks by applying the chain rule. Backpropagation calculations work "backward" from the last neural network layer to the first, updating weights individually so that loss functions are reduced over subsequent training iterations. Also called "backward propagation of errors." |
Thomas Wood, "Machine Learning Glossary," DeepAI, https://deepai.org/machine-learning-glossary-and-terms/backpropagation.
Stefania Cristina, "The Chain Rule of Calculus for Univariate and Multivariate Functions," Machine Learning Mastery: Calculus for Machine Learning (tutorial), March 16, 2022, https://machinelearningmastery.com/the-chain-rule-of-calculus-for-univariate-and-multivariate-functions/. |
Learning Methods and Techniques |
| Benchmark |
In the context of artificial intelligence,
1) a structured way of comparing the performance of different machine learning models (or hardware).
2) a widely used and publicly available dataset.
3) the highest currently achieved performance in a given task.
4) a publicly hosted machine learning challenge. |
Ramona Leenings, Nils R. Winter, Udo Dannlowski, and Tim Hahn, "Recommendations for Machine Learning Benchmarks in Neuroimaging," NeuroImage 257 (August 15, 2022): 1-9, https://doi.org/10.1016/j.neuroimage.2022.119298. |
Performance and Evaluation |
| BIG-bench |
"The Beyond the Imitation Game Benchmark (BIG-bench) is a collaborative benchmark intended to probe large language models and extrapolate their future capabilities." Created by Google researchers, BIG-bench allows anyone to submit additional information to support benchmark tasks. |
Aarohi Srivastava et al., "Beyond the Imitation Game: Quantifying and Extrapolating the Capabilities of Language Models," arXiv preprint, submitted June 9, 2022, https://arxiv.org/abs/2206.04615. |
Performance and Evaluation |
| Bilingual Evaluation Understudy (BLEU) |
In the context of artificial intelligence and natural language processing more specifically, an evaluation metric for machine translation that measures how close a text is to reference translation texts. Similar metrics include METEOR (Metric for Evaluation of Translation with Explicit Ordering) and WER (Word Error Rate). BLEU is not commonly used as of 2023 because of advances in performance evaluation. |
Benjamin Marie, "BLEU: A Misunderstood Metric from Another Age," Towards Data Science, November 4, 2022, https://towardsdatascience.com/bleu-a-misunderstood-metric-from-another-age-d434e18f1b37. |
Performance and Evaluation |
| Books3 |
A massive text corpus of more than 190,000 books in plain-text format. It was part of "The Pile" used to train large language models and has been taken down following piracy complaints. |
Alex Perry, "A Giant Online Book Collection Meta Used to Train Its AI Is Gone Over Copyright Issues," Mashable, August 18, 2023, https://mashable.com/article/books3-ai-training-dmca-takedown. |
Datasets |
| Catastrophic forgetting |
In the context of artificial intelligence and neural networks more specifically, a situation when the performance of a network on a previously trained task dramatically degrades, usually when that network is trained on a new task. |
James Kirkpatrick et al., "Overcoming Catastrophic Forgetting in Neural Networks," PNAS 114, no. 13 (March 28, 2017): 3521-3526, https://doi.org/10.1073/pnas.1611835114. |
Other |
| Chain-of-thought prompting |
A prompt engineering technique that uses specific phrasing to induce a large language model to "think step by step" or otherwise produce intermediate reasoning steps that constitute a human-interpretable explanation. Chain-of-thought prompting has been shown to improve the performance of large language models on several tasks. It is an early example of an emergent capability. |
Sunil Ramlochan, "Master Prompting Concepts: Chain of Thought Prompting," Prompt Engineering Institute, April 26, 2023, https://www.promptengineering.org/master-prompting-concepts-chain-of-thought-prompting/. |
Prompting Technique |
| Chatbot |
A computer program that simulates and processes human conversation. It can be strictly rules-based or be based upon more sophisticated information-retrieval and text-generation systems. |
"What Is a Chatbot?" Oracle Cloud Infrastructure, https://www.oracle.com/chatbots/what-is-a-chatbot/. |
Applications and Use Cases |
| Classification |
In the context of artificial intelligence, a method for assigning a correct label or class to a data input. |
"Classification in Machine Learning: An Introduction," DataCamp, updated September 2022, https://www.datacamp.com/blog/classification-machine-learning. |
Applications and Use Cases |
| Co-learning |
In the context of artificial intelligence,
1) joint training of a student and teacher encoder model using a common decoder. Co-learning can be used to reduce the size of (compress) a model.
2) use of one artificial intelligence model to train another.
3) a situation in which humans and machines adapt to and learn from one another. |
Rupak Vignesh Swaminathan, Brian King, Grant Strimel, Jasha Droppo, and Athanasios Mouchtaris, "CoDERT: Distilling Encoder Representations with Co-learning for Transducer-Based Speech Recognition," Amazon Science, 2021, https://www.amazon.science/publications/codert-distilling-encoder-representations-with-co-learning-for-transducer-based-speech-recognition.
"LLMs Are Amazing. What's Next? Agile AI," Evidentli, updated May 1, 2023, https://www.evidentli.com/news/agile-ai. |
Learning Methods and Techniques |
| Code generation |
In the context of generative artificial intelligence, a form of text generation where the text language generated is a programming language. Programs generated can be used independently of the generator program. |
https://www.ibm.com/blog/ai-code-generation/ |
Applications and Use Cases |
| Common Crawl |
A public web text dataset regularly collected since 2008. The Common Crawl corpus contains raw web page data, metadata extracts, and text extracts totaling petabytes of data. |
https://commoncrawl.org/overview |
Datasets |
| Computational costs |
In the context of artificial intelligence, the combination of
1) the cost in terms of time, kilowatts of energy, floating-point operations, or number of computations needed to perform a particular task, such as training an artificial intelligence model, and
2) the cost of a stack of computer resources, including hardware (e.g., chips) and software programs needed to train an artificial intelligence model or perform other computational tasks. |
Jai Vipra and Sarah Myers West, "Computational Power and AI," AI Now Institute, September 27, 2023, https://ainowinstitute.org/publication/policy/compute-and-ai. |
Hardware and Computational Resources |
| Constitutional artificial intelligence |
Principle-based oversight of artificial intelligence systems that uses a combination of supervised learning and reinforcement learning to align the artificial intelligence with those principles. See also "reinforcement learning from artificial intelligence feedback." |
"Constitutional AI: Harmlessness from AI Feedback," Anthropic, updated December 15, 2022, https://www.anthropic.com/index/constitutional-ai-harmlessness-from-ai-feedback. |
Learning Methods and Techniques |
| Context length |
In the context of generative artificial intelligence, the maximum length of tokens a large language model can process at a single time, including both prompt and generated responses. Increasing context length increases the amount of data that can be used to inform a response and the amount of data the model can generate without forgetting earlier parts. Sometimes also called "context window." |
"Context Length in LLMs: All You Need to Know," AGI Sphere, August 23, 2023, https://agi-sphere.com/context-length/.
"Input and Output Sizes," Anthropic, updated November 2023, https://docs.anthropic.com/claude/reference/input-and-output-sizes. |
Performance and Evaluation |
| Contrastive Language-Image Pre-training (CLIP) |
"A neural network trained on a variety of (image, text) pairs. It can be instructed in natural language to predict the most relevant text snippet, given an image, without directly optimizing for the task." CLIP was open-sourced by OpenAI in 2021 and has seen widespread industry adoption. |
OpenAI, "CLIP: Connecting Text and Images," updated January 2021, https://github.com/openai/CLIP. |
Data Handling and Processing |
| Convolutional neural network |
A type of artificial intelligence model architecture often used for image analysis and classification that is characterized by the connection of neurons in layers with at least one layer performing convolutional operations. |
Apostol Vassilev, Alina Oprea, Alie Fordyce, and Hyrum Anderson, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations, NIST Trustworthy and Responsible AI, NIST AI 100-2e2023 (Gaithersburg, MD: National Institute of Standards and Technology, January 2024), https://doi.org/10.6028/NIST.AI.100-2e2023. |
AI Model Types and Architectures |
| Data annotation |
A synonym for "annotation." |
NaN |
Data Handling and Processing |
| Data poisoning |
A form of adversarial attack on an artificial intelligence model in which an adversary gains influence over the model's training by inserting or modifying training examples. |
Apostol Vassilev, Alina Oprea, Alie Fordyce, and Hyrum Anderson, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations, NIST Trustworthy and Responsible AI, NIST AI 100-2e2023 (Gaithersburg, MD: National Institute of Standards and Technology, January 2024), https://doi.org/10.6028/NIST.AI.100-2e2023. |
Security |
| Data privacy attack |
A form of attack against an artificial intelligence model designed to gain access to sensitive information contained in training data. See "data reconstruction." |
Apostol Vassilev, Alina Oprea, Alie Fordyce, and Hyrum Anderson, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations, NIST Trustworthy and Responsible AI, NIST AI 100-2e2023 (Gaithersburg, MD: National Institute of Standards and Technology, January 2024), https://doi.org/10.6028/NIST.AI.100-2e2023. |
Security |
| Data reconstruction |
A form of "data privacy attack" designed to gain access to training data for the purpose of reconstructing sensitive information the data contain. |
Apostol Vassilev, Alina Oprea, Alie Fordyce, and Hyrum Anderson, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations, NIST Trustworthy and Responsible AI, NIST AI 100-2e2023 (Gaithersburg, MD: National Institute of Standards and Technology, January 2024), https://doi.org/10.6028/NIST.AI.100-2e2023. |
Security |
| Data wrangling |
The process of converting raw data into a computer-usable form. Also called "data cleaning," "data munging," and "data remediation." |
CDAO CTO Team |
Data Handling and Processing |
| Decoder |
A type of artificial intelligence model that reconstructs high-dimensional data from lower-dimensional representations by remapping inputs and their weights through the hidden layer of a neural network. See also "encoder." |
"Machine Learning Glossary," DeepAI, https://deepai.org/machine-learning-glossary-and-terms/autoencoder. |
AI Model Types and Architectures |
| Decoder-only model |
Any transformer-based artificial intelligence model that uses only the decoder part of an encoder-decoder model. Decoder-only models work well for text-generation tasks because their pre-training focuses on predicting the next tokens in a sequence. The GPT family of models are decoder-only models. See also "encoder-only model." |
"Decoder Models," Hugging Face, https://huggingface.co/learn/nlp-course/chapter1/6. |
AI Model Types and Architectures |
| Deep learning |
In the context of artificial intelligence, there are many definitions. A common definition is a subset of machine learning that teaches computers to process data in a way that is inspired by the neuronal structure of a mammalian brain. Deep learning neural networks, or artificial neural networks, are made of many layers of artificial neurons, which are software modules called nodes, that use mathematical calculations to process data. The layers in a deep learning neural network are the "input layers" or nodes that input data to the algorithms; "hidden layers" that process information to identify patterns; and "output layers" or nodes that give "answers" such as "yes/ no" or "cat/dog." |
"What Is Deep Learning," Amazon Web Services, https://aws.amazon.com/what-is/deep-learning. |
AI Model Types and Architectures |
| Deployment stage |
In the context of artificial intelligence, a stage in the machine learning pipeline where models are incorporated into a software environment with new, real-world data, such as an app, so that end users can interact with the model. |
"What Is Model Deployment?" Iguazio, https://www.iguazio.com/glossary/model-deployment/.
Apostol Vassilev, Alina Oprea, Alie Fordyce, and Hyrum Anderson, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations, NIST Trustworthy and Responsible AI, NIST AI 100-2e2023 (Gaithersburg, MD: National Institute of Standards and Technology, January 2024), https://doi.org/10.6028/NIST.AI.100-2e2023. |
AI Ethics and Governance |
| Diffusion model |
A type of generative artificial intelligence model wherein noise is added to input data in a step-by-step fashion (the forward diffusion process) and then the inputs are reconstructed through removal of the noise (the reverse diffusion process). A diffusion model can be modified through application of more complex data distributions to obtain new data representations. Leading image-generation models such as Midjourney, Stable Diffusion, and DALL-E are diffusion models. |
Lilian Weng, "What Are Diffusion Models?" Lil'Log, July 11, 2021, https://lilianweng.github.io/posts/2021-07-11-diffusion-models/.
Akruti Acharya, "An Introduction to Diffusion Models for Machine Learning," Encord, August 8, 2023, https://encord.com/blog/diffusion-models/. |
AI Model Types and Architectures |
| Dimensionality reduction |
In the context of artificial intelligence, a process for reducing the number of features in a dataset, whether selective (choosing to keep some features) or extractive (reducing the number of features by creating new ones). |
Joseph Rocca, "Understanding Variational Autoencoders (VAEs): Building, Step by Step, the Reasoning That Leads to VAEs," Towards Data Science, September 23, 2019, https://towardsdatascience.com/understanding-variational-autoencoders-vaes-f70510919f73. |
Applications and Use Cases |
| Directional stimulus prompting |
A method of using a smaller language model to guide a larger model for the purpose of overcoming the alignment problem by automatically incorporating keywords into the prompt so that the model generates outputs closer to the desired format and detail. |
Zekun Li et al., "Guiding Large Language Models via Directional Stimulus Prompting," arXiv preprint, submitted October 9, 2023, https://arxiv.org/pdf/2302.11520.pdf. |
Prompting Technique |
| Distributional robustness |
In the context of artificial intelligence, a characteristic of models that can provide more equitable responses over the range of possible classes, including rare or long-tail classes. Because the models are trained using different loss functions that depend on different class characteristics, they can respond appropriately to out-of-distribution cases seen during training. |
Dvir Samuel and Gal Chechik, "Distributional Robustness Loss for Long-tail Learning," arXiv preprint, submitted April 7, 2021, https://arxiv.org/abs/2104.03066. |
AI Ethics and Governance |
| Dual-use foundation model |
"An AI model that is trained on broad data; generally uses self-supervision; contains at least tens of billions of parameters; is applicable across a wide range of contexts; and that exhibits, or could be easily modified to exhibit, high levels of performance at tasks that pose a serious risk to security, national economic security, national public health or safety, or any combination of those matters, such as by: (i) substantially lowering the barrier of entry for non-experts to design, synthesize, acquire, or use chemical, biological, radiological, or nuclear (CBRN) weapons; (ii) enabling powerful offensive cyber operations through automated vulnerability discovery and exploitation against a wide range of potential targets of cyber attacks; or (iii) permitting the evasion of human control or oversight through means of deception or obfuscation. Models meet this definition even if they are provided to end users with technical safeguards that attempt to prevent users from taking advantage of the relevant unsafe capabilities." |
Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence, Sections 3(k)(i-iii), October 20, 2023, https://www.whitehouse.gov/briefing-room/presidential-actions/2023/10/30/executive-order-on-the-safe-secure-and-trustworthy-development-and-use-of-artificial-intelligence/. |
AI Model Types and Architectures |
| Embedding |
In the context of artificial intelligence, a form of data representation that carries semantic meaning by transforming objects and concepts into lists of numbers (vectors) that quantify the relationship between objects and concepts. This quantification can improve the ability of an artificial intelligence model to find relevant data relative to traditional search engines, which can only return exact matches to a query. Embeddings can be constructed for other kinds of data beside text, such as images and audio, and are typically obtained from models specifically trained for the purpose of making embeddings that capture semantic meaning well. See also "latent space." |
Rajat Tripathi, "What Are Vector Embeddings?" Pinecone, https://www.pinecone.io/learn/vector-embeddings/. |
Data Handling and Processing |
| Embedding layers |
In the context of artificial intelligence and neural networks more specifically, a type of hidden layer in a neural network that maps input information from a high-dimensional to a lower-dimensional space, allowing the network to learn more about the relationship between inputs and to process the data more efficiently. These layers are what creates embeddings. |
Enes Zvornicanin, "What Are Embedding Layers in Neural Networks?" Baeldung, updated May 24, 2023, https://www.baeldung.com/cs/neural-nets-embedding-layers. |
Neural Network Components and Functions |
| Emergent capabilities (aka emergent properties) |
In the context of generative artificial intelligence, many properties can fall under this concept, including unexpected output that can be positive, negative, or neutral, as demonstrated by generative artificial intelligence models on tasks that they were not specifically trained for. |
Rishi Bommasani, "On the Opportunities and Risks of Foundation Models," arXiv preprint, submitted July 12, 2022, https://arxiv.org/pdf/2108.07258.pdf. |
Performance and Evaluation |
| Encoder |
A type of artificial intelligence model that compresses high-dimensional data, such as text, into a lower dimension, such as numbers. An encoder passes the input data into the hidden layers of a neural network. See also "decoder." |
"Machine Learning Glossary," DeepAI, https://deepai.org/machine-learning-glossary-and-terms/autoencoder. |
AI Model Types and Architectures |
| Encoder-decoder model |
A type of artificial intelligence model suited for generative tasks. It relies on encoders to encode input data into manipulable vectors and on decoders to manipulate the resulting vectors into human-interpretable data, such as in tasks like translation and question answering. |
Daniel Nelson, "What Is an Autoencoder?" Unite.AI, updated September 20, 2020, https://www.unite.ai/what-is-an-autoencoder/. |
AI Model Types and Architectures |
| Encoder-only model |
A type of artificial intelligence model that uses only the encoder part of a transformer model. Encoder-only models are useful for classification and extraction tasks because their training focuses on reconstructing a sequence of tokens. The BERT family of models are encoder-only models. See also "decoder-only model." |
"Encoder Models," Hugging Face, https://huggingface.co/learn/nlp-course/chapter1/5?fw=pt. |
AI Model Types and Architectures |
| End-of-sequence token |
In the context of generative artificial intelligence, a string of characters a model generates that lets the model know it can stop generating text. Also called "stop sequence." See also "start-of-sequence token." |
Ashton Zhang et al., "Sequence to Sequence Learning," in Dive Into Deep Learning, https://classic.d2l.ai/chapter_recurrent-modern/seq2seq.html. |
Natural Language Processing |
| Equity |
A principle that requires that an artificial intelligence system, including human developers, take deliberate steps to minimize the presence of unintended bias in inputs and the effects of unintended socially relevant bias from the outputs of artificial intelligence models. |
Department of Defense, "Responsible Artificial Intelligence Strategy and Implementation Pathway" (Washington, DC: USD(AT&L), June 2022), https://media.defense.gov/2022/Jun/22/2003022604/-1/-1/0/Department-of-Defense-Responsible-Artificial-Intelligence-Strategy-and-Implementation-Pathway.PDF. |
AI Ethics and Governance |
| Explainability |
"A characteristic of an AI system in which there is provision of accompanying evidence or reasons for system output in a manner that is meaningful or understandable to individual users (as well as developers and auditors) and reflects the system's process for generating the output (e.g., what alternatives were considered, but not proposed, and why not)." |
Department of Defense, "Responsible Artificial Intelligence Strategy and Implementation Pathway" (Washington, DC: USD(AT&L), June 2022), https://media.defense.gov/2022/Jun/22/2003022604/-1/-1/0/Department-of-Defense-Responsible-Artificial-Intelligence-Strategy-and-Implementation-Pathway.PDF. |
AI Ethics and Governance |
| Extraction |
A type of adversarial attack in which the attacker seeks to access the training data of a generative artificial intelligence model. |
Apostol Vassilev, Alina Oprea, Alie Fordyce, and Hyrum Anderson, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations, NIST Trustworthy and Responsible AI, NIST AI 100-2e2023 (Gaithersburg, MD: National Institute of Standards and Technology, January 2024), https://doi.org/10.6028/NIST.AI.100-2e2023. |
Security |
| Feed-forward neural network |
In the context of artificial intelligence and neural networks more specifically, a form of neural network model in which there is a unidirectional flow of information from input to hidden layers to output nodes without recurrent or looping cycles. |
Apostol Vassilev, Alina Oprea, Alie Fordyce, and Hyrum Anderson, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations, NIST Trustworthy and Responsible AI, NIST AI 100-2e2023 (Gaithersburg, MD: National Institute of Standards and Technology, January 2024), https://doi.org/10.6028/NIST.AI.100-2e2023. |
AI Model Types and Architectures |
| Fine-tuning |
In the context of generative artificial intelligence, any of a range of techniques used to modify a pre-trained model's weights such that it returns results more appropriate to a specific domain. See also "reinforcement learning from artificial intelligence feedback," "reinforcement learning from human feedback," and "supervised fine-tuning." |
Chip Huyen, "RLHF: Reinforcement Learning from Human Feedback," updated May 2, 2023, https://huyenchip.com/2023/05/02/rlhf.html. |
Learning Methods and Techniques |
| Floating-point operation (FLOP) |
"Any mathematical operation or assignment involving floating-point numbers, which are a subset of the real numbers typically represented on computers by an integer of fixed precision scaled by an integer exponent of a fixed base." |
Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence, Section 3(m), October 20, 2023, https://www.whitehouse.gov/briefing-room/presidential-actions/2023/10/30/executive-order-on-the-safe-secure-and-trustworthy-development-and-use-of-artificial-intelligence/. |
Hardware and Computational Resources |
| Foundation model |
In the context of generative artificial intelligence, a type of artificial intelligence model that is trained on broad data (generally using self-supervision at scale) and that can be adapted (e.g., fine-tuned) to a wide range of tasks. |
Rishi Bommasani, "On the Opportunities and Risks of Foundation Models," arXiv preprint, submitted August 16, 2021, https://arxiv.org/abs/2108.07258. |
AI Model Types and Architectures |
| Fully connected layer |
In the context of artificial intelligence and neural networks more specifically, a type of neural network layer that maps a connection of every input neuron to every output neuron. |
"Optimizing Linear/Fully-Connected Layers: User's Guide," Nvidia Docs Hub, https://docs.nvidia.com/deeplearning/performance/dl-performance-fully-connected/index.html#fullyconnected-layer. |
Neural Network Components and Functions |
| Functional attack |
A form of adversarial attack that exploits characteristics of the types of data (e.g., images) used by a model to alter the model's behavior. |
Apostol Vassilev, Alina Oprea, Alie Fordyce, and Hyrum Anderson, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations, NIST Trustworthy and Responsible AI, NIST AI 100-2e2023 (Gaithersburg, MD: National Institute of Standards and Technology, January 2024), https://doi.org/10.6028/NIST.AI.100-2e2023. |
Security |
| Generative adversarial network (GAN) |
A generative artificial intelligence approach wherein a pair of supervised learning models are trained collaboratively "based on a game theoretic scenario in which the generator network must compete against an adversary. The generator network directly produces samples. Its adversary, the discriminator network, attempts to distinguish between samples drawn from the training data and samples drawn from the generator." GANS were state of the art for image generation before diffusion models gained popularity. |
Jason Brownlee, "A Gentle Introduction to Generative Adversarial Networks (GANs),"
updated July 19, 2019, Machine Learning Mastery, https://machinelearningmastery.com/what-are-generative-adversarial-networks-gans/. |
AI Model Types and Architectures |
| Generative artificial intelligence |
A generic term for any AI system that generates content such as text, imagery, or other modalities. |
Helen Toner, "What Are Generative AI, Large Language Models, and Foundation Models?" Center for Security and Emerging Technology, updated May 12, 2023, https://cset.georgetown.edu/article/what-are-generative-ai-large-language-models-and-foundation-models/. |
AI Model Types and Architectures |
| Generative penalty |
A set of techniques used to alter outputs of generative models, such as by assigning a frequency penalty when a generated token appears more frequently than preferred or a presence penalty when a specific generated token is not to be included in outputs at all. |
"Concepts for Generative AI," Oracle, updated September 18 2023, https://docs.oracle.com/en-us/iaas/Content/generative-ai/concepts.htm. |
Sampling Methods |
| Governable |
In the context of artificial intelligence, a principle describing artificial intelligence capabilities that fulfill their intended functions while possessing both the ability to detect and avoid unintended consequences and the ability to disengage or deactivate deployed systems that demonstrate unintended behavior. |
Department of Defense, "Responsible Artificial Intelligence Strategy and Implementation Pathway" (Washington, DC: USD(AT&L), June 2022), https://media.defense.gov/2022/Jun/22/2003022604/-1/-1/0/Department-of-Defense-Responsible-Artificial-Intelligence-Strategy-and-Implementation-Pathway.PDF. |
AI Ethics and Governance |
| Graphics processing unit (GPU) |
A specialized processor that can process large amounts of data simultaneously. Without GPUs or a comparable application-specific integrated circuit, it is not computationally cost effective to train large language models or provide inference at scale. As of December 2023, NVIDIA is the market leader in developing GPUs for generative artificial intelligence training and inference. |
"What Is a GPU?" Intel, https://www.intel.com/content/www/us/en/products/docs/processors/what-is-a-gpu.html. |
Hardware and Computational Resources |
| Greedy decoding |
A configuration parameter for a large language model that tells the model to always generate the token with the highest probability of being linked to the previous token. This method tends to produce less natural sentences than "top-k sampling" or "top-p sampling." |
Fabio Chiusano, "Two Minutes NLP - Most Used Decoding Methods for Language Models," NLPlanet, updated January 28, 2022, https://medium.com/nlplanet/two-minutes-nlp-most-used-decoding-methods-for-language-models-9d44b2375612. |
Sampling Methods |
| Hallucination |
In the context of generative artificial intelligence, any of numerous kinds and degrees of incidents in which a large language model generates an inaccurate but plausible-sounding term or phrase in response to a prompt based upon the model's perception of patterns in the data. |
"What Are AI Hallucinations?" IBM, https://www.ibm.com/topics/ai-hallucinations. |
Performance and Evaluation |
| Holistic Evaluation of Language Models (HELM) |
A benchmark for large language model transparency that uses 6 core scenarios (question answering, information retrieval, summarization, sentiment analysis, toxicity detection, and text classification); 7 metrics (accuracy, calibration, robustness, fairness, bias, toxicity, and efficiency); and 7 targeted evaluations (language, knowledge, reasoning, memorization and copyright, disinformation, bias, and toxicity) to evaluate 30 common large language models. HELM is maintained by its creators, the Stanford University Institute for Human-Centered Artificial Intelligence (HAI). |
"A Holistic Framework for Evaluating Foundation Models," Center for Research on Foundation Models, https://crfm.stanford.edu/helm/latest/. |
Performance and Evaluation |
| Human-system integration |
An approach applied to system development and integration that emphasizes interdisciplinary evaluation to ensure that human performance is optimized to increase total system performance and minimize total system ownership costs. DoD 5000.05 E7 outlines the many components of human-systems-interaction, including human factors engineering and cognitive engineering, and the negative characteristics of human-machine interactions which must be minimized, including excessive cognitive, physical, or sensory skill demands or health hazards |
Department of Defense, "Responsible Artificial Intelligence Strategy and Implementation Pathway" (Washington, DC: USD(AT&L), June 2022), https://media.defense.gov/2022/Jun/22/2003022604/-1/-1/0/Department-of-Defense-Responsible-Artificial-Intelligence-Strategy-and-Implementation-Pathway.PDF.
Department of Defense, "DoD Directive 5000.01: The Defense Acquisition System" (Washington, DC: USD(A&S), September 2020), https://www.esd.whs.mil/Portals/54/Documents/DD/issuances/dodd/500001p.pdf |
AI Ethics and Governance |
| Hyperparameter |
In the context of artificial intelligence, any top-level, externally configurable variable for machine learning model training that is supplied by a developer and not learned from data. For large language models, a hyperparameter may include the number of attention heads, the context length, and other configuration variables common to training any deep learning model, such as learning rate, batch size, and number of training iterations (epochs). Without knowing about these configurations, it may be nearly impossible to reproduce a trained model. |
"What Is Hyperparameter Tuning?" Amazon Web Services, https://aws.amazon.com/what-is/hyperparameter-tuning.
Tales Matos, "Tuning Parameters to Train LLMs," Medium, updated July 24, 2023, https://medium.com/@rtales/tuning-parameters-to-train-llms-large-language-models-8861bbc11971. |
Learning Methods and Techniques |
| Image segmentation |
A process of dividing a digital image into regions to reduce the complexity of the image and locate objects and boundaries by assigning a label to every pixel in the image. See also "instance segmentation." |
CDAO CTO Team |
Applications and Use Cases |
| In-context learning |
In the context of generative artificial intelligence, a specific method of prompt engineering wherein demonstrations of the task are provided to the model as part of the prompt (in natural language). See "N-shot prompting." |
"What Is In Context Learning?" Hopsworks, https://www.hopsworks.ai/dictionary/in-context-learning-icl. |
Prompting Technique |
| Inference |
In the context of artificial intelligence, "the process of running data points into a machine learning model to calculate an output such as a single numerical score. This process is also referred to as operationalizing a machine learning model or putting a machine learning model into production." |
"Model Inference Overview," Google Cloud Big Query, https://cloud.google.com/bigquery/docs/inference-overview. |
Other |
| Information extraction |
In the context of artificial intelligence, any of multiple computer tasks, such as
1) the production of structured data from unstructured or semi-structured documents or other data modalities. An example is creating a table of companies and their market cap from the text of a market research report.
2) the pulling of structured data from unstructured text or images. |
"Information Extraction," SnorkelAI, https://snorkel.ai/information-extraction/. |
Applications and Use Cases |
| Instance segmentation |
A process of dividing a digital image into distinct instances of objects when multiples of those objects exist in an image. See also "image segmentation." |
CDAO CTO Team |
Applications and Use Cases |
| Integer operation |
Any mathematical operation on whole numbers. |
Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence, Section 3(r), October 20, 2023, https://www.whitehouse.gov/briefing-room/presidential-actions/2023/10/30/executive-order-on-the-safe-secure-and-trustworthy-development-and-use-of-artificial-intelligence/. |
Hardware and Computational Resources |
| Jailbreak |
In the context of generative artificial intelligence, any of numerous methods, such as the use of engineered prompts, to cause a model to override its alignment safeguards. For example, one could get a large language model to provide bomb-making instructions by having it first pretend it is writing a screenplay for a movie. |
Patrick Chao et al., "Jailbreaking Black Box Large Language Models in Twenty Queries," arXiv preprint, submitted October 12, 2023, https://arxiv.org/abs/2310.08419. |
Security |
| Labeling |
A synonym for "annotation." |
CDAO CTO Team |
Data Handling and Processing |
| Language model |
In the context of artificial intelligence, any deep-learning technique or statistical technique by which a model learns from patterns in natural language in order to perform text-classification or text-generation functions. |
Daniel Atherton, Reva Schwartz, Peter C. Fontana, and Patrick Hall, The Language of Trustworthy AI: An In-Depth Glossary of Terms (Gaithersburg, MD: National Institute of Standards and Technology, 2023), https://doi.org/10.6028/NIST.AI.100-3. |
AI Model Types and Architectures |
| Large language model (LLM) |
In the context of artificial intelligence, a class of language models that use deep-learning algorithms and are pre-trained on extremely large textual datasets that can be multiple terabytes in size. LLMs can be classed into two types: generative or discriminatory. Generative LLMs like GPT-4 are models that output text, such as the answer to a question or an essay on a specific topic. Discriminatory LLMs like BERT are supervised learning models that usually focus on classifying text, such as determining whether a text was made by a human or an artificial intelligence. As of December 2023, state-of-the-art LLMs also included Llama-2. |
Daniel Atherton, Reva Schwartz, Peter C. Fontana, and Patrick Hall, The Language of Trustworthy AI: An In-Depth Glossary of Terms (Gaithersburg, MD: National Institute of Standards and Technology, 2023), https://doi.org/10.6028/NIST.AI.100-3. |
AI Model Types and Architectures |
| Latent space |
In the context of artificial intelligence, there are many definitions. Generally, however, this term refers to the abstract multi-dimensional space that encodes a meaningful internal representation of externally observed events. See also "embedding." |
Panagiotis Antoniadis, "Latent Space in Deep Learning," Baeldung, updated June 11, 2023, https://www.baeldung.com/cs/dl-latent-space. |
Data Handling and Processing |
| Loss function |
A mathematical process that quantifies the error margin between a model's prediction and the actual target value. The loss function is a measurable way to gauge the performance and accuracy of a machine learning model. |
"Loss Functions in Machine Learning Explained," Datacamp, updated November 2023, https://www.datacamp.com/tutorial/loss-function-in-machine-learning. |
Other |
| Lossless encoding |
In the context of artificial intelligence, an encoding technique defined by the preservation of information or the prevention of loss of information during the encoding process. See also "lossy encoding." |
Joseph Rocca, "Understanding Variational Autoencoders (VAEs): Building, Step by Step, the Reasoning That Leads to VAEs," Towards Data Science, September 23, 2019, https://towardsdatascience.com/understanding-variational-autoencoders-vaes-f70510919f73. |
Data Handling and Processing |
| Lossy encoding |
In the context of artificial intelligence, an encoding technique defined by the loss of some information during the dimensionality reduction that cannot be retrieved during decoding. Large language models are often analogized as lossy encodings of their training data. See also "lossless encoding." |
Joseph Rocca, "Understanding Variational Autoencoders (VAEs): Building, Step by Step, the Reasoning That Leads to VAEs," Towards Data Science, September 23, 2019, https://towardsdatascience.com/understanding-variational-autoencoders-vaes-f70510919f73. |
Data Handling and Processing |
| Low-Rank Adaptation of Large Language Models (LoRA) |
A fine-tuning technique that freezes the pre-trained weights of a model so the model instead learns a lower-dimensional weight change matrix. In practical terms, LoRA vastly reduces the computational requirements to fine-tune a large language model or a diffusion model since some weights are not being updated. |
Edward J. Hu et al., "LoRA: Low-Rank Adaptation of Large Language Models," arXiv preprint, submitted June 17, 2021, https://arxiv.org/abs/2106.09685.
Sebastian Raschka, "Parameter-Efficient LLM Finetuning With Low-Rank Adaptation (LoRA)," AI Magazine, updated April 26, 2023, https://sebastianraschka.com/blog/2023/llm-finetuning-lora.html. |
Learning Methods and Techniques |
| MAMBA |
A state-space model architecture released in December 2023 that has been shown to exhibit longer context lengths than transformer models with lower computational costs and comparable benchmark performance. |
Albert Gu and Tri Dao, "Mamba: Linear-Time Sequence Modeling with Selective State Spaces," arXiv preprint, submitted December 1, 2023, https://arxiv.org/abs/2312.00752. |
AI Model Types and Architectures |
| Manual labeling |
A method of annotating data in which a human labels each data instance. See also "programmatic labeling." |
"What Is Programmatic AI?" Snorkel AI, https://snorkel.ai/programmatic-labeling/. |
Data Handling and Processing |
| Mask filling |
In the context of artificial intelligence, a method for testing the ability of a language model to complete a sequence of tokens after a token or tokens have been obscured to the model. |
"Fill-Mask," Hugging Face, https://huggingface.co/tasks/fill-mask. |
Learning Methods and Techniques |
| Massive Multitask Language Understanding (MMLU) |
A benchmark designed to "measure knowledge acquired during pretraining" in zero shot and few shot settings." The benchmark data cover 57 subjects across science, technology, engineering, and mathematics (STEM); the humanities; the social sciences; and more. The data range in difficulty from an elementary level to an advanced professional level, and they test both world knowledge and problem-solving ability. Subjects range from traditional areas, such as mathematics and history, to more specialized areas, such as law and ethics. |
Dan Hendrycks, "Measuring Massive Multitask Language Understanding," arXiv preprint, submitted January 12, 2021, https://arxiv.org/pdf/2009.03300.pdf. |
Performance and Evaluation |
| MATH |
In the context of artificial intelligence, a benchmark of 12,500 competition mathematics problems designed to test the performance of an artificial intelligence model on mathematical problem-solving. |
Dan Hendrycks et al., "Measuring Mathematical Problem Solving With the MATH Dataset," arXiv preprint, submitted March 5, 2021, https://arxiv.org/abs/2103.03874. |
Performance and Evaluation |
| Model architecture |
In the context of artificial intelligence, the choice of a machine learning algorithm along with the underlying structure or design of the machine learning model, such as layers of interconnected nodes or neurons, where each layer of the model performs a specific function, such as data pre-processing, feature extraction, or prediction, depending on the type of problem being solved, the size and complexity of the dataset, and the available computing resources. |
"Model Architecture," Hopsworks, https://www.hopsworks.ai/dictionary/model-architecture. |
AI Model Types and Architectures |
| Model hub |
A website that hosts trained model weights and includes context on their creation and intended use. Model hubs aid model discovery and reuse. Examples of model hubs include Huggingface's Model Hub, PyTorch Hub, Tensorflow Hub, Model Zoo, and ONNX Model Zoo. Also called "model zoo." |
"Model Zoo," SaturnCloud, https://saturncloud.io/glossary/model-zoo/. |
Tools and Platforms |
| Model theft |
In the context of artificial intelligence, one of numerous types of adversarial behavior of humans or artificial intelligence models designed to gain knowledge of, access to, or use of others' proprietary models. |
"OWASP Top 10 for LLM Applications: Version 1.1," Open Worldwide Application Security Project, October 16, 2023, https://owasp.org/www-project-top-10-for-large-language-model-applications/assets/PDF/OWASP-Top-10-for-LLMs-2023-v1_1.pdf. |
Security |
| Model weight |
In the context of artificial intelligence, "a numerical parameter within an AI model that helps determine the model's outputs in response to inputs." |
Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence, Section 3(u), October 20, 2023, https://www.whitehouse.gov/briefing-room/presidential-actions/2023/10/30/executive-order-on-the-safe-secure-and-trustworthy-development-and-use-of-artificial-intelligence/. |
Hardware and Computational Resources |
| Multi-modal model |
A type of generative artificial intelligence model characterized by its ability to process and analyze data of multiple types (e.g., text, images, audio). |
Nate Rosidi, "Unlocking the Power of Multimodal Learning: Techniques, Challenges, and Applications," Kdnuggets, updated March 27, 2023, https://www.kdnuggets.com/2023/03/multimodal-models-explained.html. |
AI Model Types and Architectures |
| N-shot prompting |
Any prompting technique that specifies the amount of context that must be given to a large language model before prompting it. See "in-context learning." |
"Few-Shot Prompting," Prompt Engineering Guide, updated January 12, 2024, https://www.promptingguide.ai/techniques/fewshot. |
Prompting Technique |
| Named entity recognition (NER) |
In the context of artificial intelligence, a task that a model performs to find which parts of the input text correspond to entities such as persons, locations, or organizations. |
"Transformers: What Can They Do?" Hugging Face, https://huggingface.co/learn/nlp-course/chapter1/3?fw=pt. |
Applications and Use Cases |
| Neural network |
In the context of artificial intelligence, "a method that teaches computers to process data in a way that is inspired by the human brain. It is a type of machine learning process, called deep learning, that uses interconnected nodes or neurons in a layered structure that resembles the human brain. It creates an adaptive system that computers use to learn from their mistakes and improve continuously." |
"What Is a Neural Network?" Amazon Web Services, https://aws.amazon.com/what-is/neural-network/. |
AI Model Types and Architectures |
| Off-label use |
In the context of artificial intelligence,
1) the use of data from one artificial intelligence task for another artificial intelligence task without acknowledgment by the user of the context or provenance of the data.
2) the use of a large language model or other model to perform functions that were not explicitly intended or permitted. |
Efrat Shimron, Jonathan I. Tamir, Ke Wang, and Michael Lustig, "Implicit Data Crimes: Machine Learning Bias Arising from Misuse of Public Data," Proceedings of the National Academy of Sciences 119, no. 13 (February 1, 2022): 1-11, https://doi.org/10.1073/pnas.211720311. |
AI Ethics and Governance |
| Output probability |
A score between 0 and 1 that a model estimates as the probability of a given answer, such as a class label. |
"Logistic Regression: Calculating a Probability," Google Machine Learning Education, https://developers.google.com/machine-learning/crash-course/logistic-regression/calculating-a-probability. |
Performance and Evaluation |
| Parameter |
In the context of artificial intelligence, the values, such as weights and biases in a neural network, that an algorithm learns from the data and updates as it is trained. The more parameters a model has, the more computational costs are associated with training it and conducting inference, but the model's performance also may be better because of scaling. By releasing a model's parameters, a developer can allow anyone to use that model for inference on any system that meets the computational requirements. See also "model weight." |
Jason Brownlee, "What Is the Difference Between a Parameter and a Hyperparameter?" updated June 17, 2019, Machine Learning Mastery, https://machinelearningmastery.com/difference-between-a-parameter-and-a-hyperparameter/. |
Hardware and Computational Resources |
| Plugin |
In the context of artificial intelligence, a software extension that enables users to perform tasks using large language models that are not tasks the models themselves perform. Examples of plugins include calculators, web browsers, and services that allow you to read from and write to a database. |
"OWASP Top 10 for LLM Applications: Version 1.1," Open Worldwide Application Security Project, October 16, 2023, https://owasp.org/www-project-top-10-for-large-language-model-applications/assets/PDF/OWASP-Top-10-for-LLMs-2023-v1_1.pdf. |
Tools and Platforms |
| Policy |
In the context of artificial intelligence and reinforcement learning more specifically, the guidance given to a model to direct its behavior. Policies may be functions or search processes that determine the behavior of reinforcement learning agents. |
"Elements of Reinforcement Learning," in Richard S. Sutton and Andrew G. Barto, Reinforcement Learning: An Introduction (Cambridge, MA: MIT Press, 1998), http://incompleteideas.net/book/ebook/node9.html. |
Learning Methods and Techniques |
| Positional encoding |
In the context of artificial intelligence and embedding more specifically, the assignment of a numeric value to a token in an input sequence, such as a sentence, that captures the order of tokens in the sequence. |
Jason Brownlee, "What Is the Difference Between a Parameter and a Hyperparameter?" updated June 17, 2019, Machine Learning Mastery, https://machinelearningmastery.com/a-gentle-introduction-to-positional-encoding-in-transformer-models-part-1/. |
Natural Language Processing |
| Pre-training |
The process of training a foundational artificial intelligence model to perform a task using large amounts of non-labeled data, such as the Common Crawl Corpus. For large language models, the pre-training task is to predict the next token given some existing sequence. Pre-trained models can be trained to perform other specific tasks through "fine-tuning." |
Angie Lee, "What Is a Pretrained AI Model?" NVIDIA, updated December 8, 2022, https://blogs.nvidia.com/blog/what-is-a-pretrained-ai-model/. |
Learning Methods and Techniques |
| Privacy |
A principle for preserving trustworthiness in systems that describes the sense that individuals and communities of people have, whether real or perceived, of controlling the information that others know about them. |
Anita L. Allen, "Privacy-as-Data Control: Conceptual, Practical, and Moral Limits of the Paradigm," Connecticut Law Review 32, no. 3 (Spring 2000): 861-875. |
AI Ethics and Governance |
| Privacy engineering |
A suite of methods, including controls on data ingest, transformation, and presentation of model outputs, that are explicitly chosen to reduce the risk of harm to persons' or communities' perception of the privacy of information about them. |
Sean Brooks et al., An Introduction to Privacy Engineering and Risk Management in Federal Systems, NISTIR 8062 (Gaithersburg, MD: National Institute of Standards and Technology, January 2017), https://nvlpubs.nist.gov/nistpubs/ir/2017/NIST.IR.8062.pdf.
https://doi.org/10.6028/NIST.IR.8062 |
AI Ethics and Governance |
| Privacy-enhancing technology (PET) |
"Any software or hardware solution, technical process, technique, or other technological means of mitigating privacy risks arising from data processing, including by enhancing predictability, manageability, disassociability, storage, security, and confidentiality. These technological means may include secure multiparty computation, homomorphic encryption, zero-knowledge proofs, federated learning, secure enclaves, differential privacy, and synthetic-data-generation tools." Also called "privacy-preserving technology." |
Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence, Section 3(z), October 20, 2023, https://www.whitehouse.gov/briefing-room/presidential-actions/2023/10/30/executive-order-on-the-safe-secure-and-trustworthy-development-and-use-of-artificial-intelligence/. |
AI Ethics and Governance |
| Programmatic labeling |
In the context of artificial intelligence, a method of annotating data in which humans write a function to an algorithmic system that can automatically annotate or reannotate all data instances with specifically defined data features. See also "manual labeling." |
"What Is Programmatic AI?" Snorkel AI, https://snorkel.ai/programmatic-labeling/. |
Data Handling and Processing |
| Prompt engineering |
The art of crafting the optimal textual input to elicit desirable outputs from a generative model. |
"What Is Prompt Engineering?" Amazon Web Services, https://aws.amazon.com/what-is/prompt-engineering/. |
Prompting Technique |
| Prompt extraction |
A form of adversarial attack in which the malicious user seeks to obtain prompts or other hidden information from a large language model. |
Apostol Vassilev, Alina Oprea, Alie Fordyce, and Hyrum Anderson, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations, NIST Trustworthy and Responsible AI, NIST AI 100-2e2023 (Gaithersburg, MD: National Institute of Standards and Technology, January 2024), https://doi.org/10.6028/NIST.AI.100-2e2023. |
Security |
| Prompt injection |
A form of adversarial behavior wherein a user crafts inputs that manipulate a large language model to perform unintended actions. Direct injections overwrite system prompts, while indirect injections manipulate inputs from external sources. |
"OWASP Top 10 for LLM Applications: Version 1.1," Open Worldwide Application Security Project, October 16, 2023, https://owasp.org/www-project-top-10-for-large-language-model-applications/assets/PDF/OWASP-Top-10-for-LLMs-2023-v1_1.pdf. |
Security |
| Quantization |
In the context of machine learning, the reduction of computational requirements for model inference by lowering the mathematical precision of the model's weights. The lower the final precision, the lower the computational requirements but also the higher the chance that rounding errors will negatively impact the model's performance. |
"Quantization," HuggingFace, https://huggingface.co/docs/optimum/concept_guides/quantization. |
Data Handling and Processing |
| Question answering |
A capability of generative artificial intelligence models, particularly the primary skill of chatbots that answer user queries. Large language models can demonstrate this capability through next-token generation, information retrieval and generation, or simple information retrieval from a bank of established question-and-answer pairs. |
"Question Answering," Hugging Face, https://huggingface.co/docs/transformers/main/tasks/question_answering. |
Applications and Use Cases |
| Question/answer pair (QA) |
A type of dataset used to train large language models to answer questions based upon labeled pairs of questions and answers. |
"Large Question Answering Datasets," GitHub, https://github.com/ad-freiburg/large-qa-datasets. |
Datasets |
| RAG (retrieval augmented generation) |
In the context of generative artificial intelligence, a method or framework for improving the quality and trustworthiness of large language model outputs by grounding them in external data. RAG systems often function by identifying relevant pieces of information from a database or search index, which are then combined with the user's prompt before the final output is generated. |
"What Is Retrieval-Augmented Generation?" IBM, https://research.ibm.com/blog/retrieval-augmented-generation-RAG. |
AI Model Types and Architectures |
| ReAct prompting |
A prompt engineering paradigm that uses a large language model to generate both reasoning traces and task-specific actions. ReAct prompting allows the system to create, maintain, and adjust action plans while interacting with external tools, like web browsers or databases. |
"ReAct Prompting," Prompt Engineering Guide, DAIR.AI, https://www.promptingguide.ai/techniques/react. |
Prompting Technique |
| Reasoning |
In the context of generative artificial intelligence, one of numerous capabilities of an artificial intelligence model that mimics the cognitive function of combining world models, common sense, and chain-of-thought to achieve complex cognitive tasks, including self-consistency in answers. Reasoning via planning (RAP) repurposes a large language model as both a world model and a reasoning agent, and it incorporates a principled planning algorithm (based on Monto Carlo Tree Search) for strategic exploration in the vast reasoning space of a large language model. |
Shibo Hao, Yi Gu, Haodi Ma, Joshua Jiahua Hong, Zhen Wang, Daisy Zhe Wang, and Zhiting Hu, "Reasoning with Language Model Is Planning with World Model," arXiv preprint, submitted May 23, 2023, https://arxiv.org/abs/2305.14992. |
Other |
| Recall-Oriented Understudy for Gisting Evaluation (ROUGE) |
"A set of metrics and a software package used for evaluating automatic summarization and machine translation software in natural language processing. The metrics compare an automatically produced summary or translation against a reference or a set of references (human-produced) summary or translation." ROUGE is not commonly used as of 2023 because more sophisticated measures of text generation now exist. |
"Metric Card for ROUGE," Hugging Face, https://huggingface.co/spaces/evaluate-metric/rouge. |
Performance and Evaluation |
| Red teaming |
A term borrowed from cybersecurity to mean an exercise wherein a team emulates an adversary's attack against a system so that another team emulating the systems' defenders learns to repel or mitigate harms from adversarial attacks. |
Committee on National Security Systems (CNSS), CNSS Glossary (Fort Meade, MD: CNSS, 2015), https://rmf.org/wp-content/uploads/2017/10/CNSSI-4009.pdf. |
Performance and Evaluation |
| Reinforcement learning |
In the context of artificial intelligence, one of the major forms of machine learning alongside supervised learning and unsupervised learning. In reinforcement learning the programmer instructs an agent to learn how to conduct actions to maximize a cumulative reward metric. Reinforcement learning algorithms can be either policy-free or policy-based. Policy-based agents can learn by making predictions about the consequences of actions, while policy-free agents learn by exploring and exploiting the environment. |
"Model-Free vs. Model-Based Reinforcement Learning," Baeldung, https://www.baeldung.com/cs/ai-model-free-vs-model-based. |
Learning Methods and Techniques |
| Reinforcement learning from artificial intelligence feedback (RLAIF) |
A novel version of "reinforcement learning from human feedback" wherein the reward model is created using data labeled by other large language models instead of data labeled by humans. If the large language model generating the data labels is guided by pre-defined human preferences, the process is called "constitutional artificial intelligence" after the company Anthropic published a paper that popularized the idea. See also "reinforcement learning from human feedback." |
Harrison Lee et al., "RLAIF: Scaling Reinforcement Learning from Human Feedback with AI Feedback," arXiv preprint, submitted September 1, 2023, https://arxiv.org/abs/2309.00267. |
Learning Methods and Techniques |
| Reinforcement learning from human feedback (RLHF) |
A method to fine-tune a large language model wherein humans label the goodness of generated outputs to train a reward model that the large language model's weights are then adjusted to maximize. RLHF is the method that led to the novel success of GPT-3.5 (the original large language model used in ChatGPT) over InstructGPT, which was only subject to "supervised fine-tuning." See also "reinforcement learning from artificial intelligence feedback." |
Chip Huyen, "RLHF: Reinforcement Learning from Human Feedback," updated May 2, 2023, https://huyenchip.com/2023/05/02/rlhf.html. |
Learning Methods and Techniques |
| Reliable |
A principle for ensuring the trustworthiness of artificial intelligence. Reliable artificial intelligence systems have explicit and well-defined uses whose safety, security, and effectiveness is subject to rigorous and routine testing and assurance throughout the model lifecycle. |
Department of Defense, "Responsible Artificial Intelligence Strategy and Implementation Pathway" (Washington, DC: USD(AT&L), June 2022), https://media.defense.gov/2022/Jun/22/2003022604/-1/-1/0/Department-of-Defense-Responsible-Artificial-Intelligence-Strategy-and-Implementation-Pathway.PDF. |
AI Ethics and Governance |
| Responsible |
A principle for ensuring the trustworthiness of artificial intelligence. Responsible artificial intelligence requires that personnel involved in any phase of the artificial intelligence lifecycle exercise appropriate levels of judgment and care to guide the development, deployment, and use of artificial intelligence capabilities. |
Department of Defense, "Responsible Artificial Intelligence Strategy and Implementation Pathway" (Washington, DC: USD(AT&L), June 2022), https://media.defense.gov/2022/Jun/22/2003022604/-1/-1/0/Department-of-Defense-Responsible-Artificial-Intelligence-Strategy-and-Implementation-Pathway.PDF. |
AI Ethics and Governance |
| Reward function |
In the context of artificial intelligence and reinforcement learning more specifically, a type of mathematical function that maps state-action pairs in a reinforcement learning algorithm to a reward number that corresponds to the desirability of that state according to the value of the short-term payoff, not necessarily the end goal of reward maximization. See also "value function." |
"Elements of Reinforcement Learning," in Richard S. Sutton and Andrew G. Barto, Reinforcement Learning: An Introduction (Cambridge, MA: MIT Press, 1998), http://incompleteideas.net/book/ebook/node9.html. |
Learning Methods and Techniques |
| Robust artificial intelligence |
A principle for ensuring the trustworthiness of artificial intelligence. Robust artificial intelligence is resilient to significant changes to input data in real-world settings and effectively repels adversarial attacks. |
Department of Defense, "Responsible Artificial Intelligence Strategy and Implementation Pathway" (Washington, DC: USD(AT&L), June 2022), https://media.defense.gov/2022/Jun/22/2003022604/-1/-1/0/Department-of-Defense-Responsible-Artificial-Intelligence-Strategy-and-Implementation-Pathway.PDF. |
AI Ethics and Governance |
| Scaling |
In the context of artificial intelligence,
1) a method to coerce a capability.
2) the capability of a model to increase or decrease the computational resources required to execute a varying volume of tasks, processes, or services. |
Daniel Atherton, Reva Schwartz, Peter C. Fontana, and Patrick Hall, The Language of Trustworthy AI: An In-Depth Glossary of Terms (Gaithersburg, MD: National Institute of Standards and Technology, 2023), https://doi.org/10.6028/NIST.AI.100-3. |
Hardware and Computational Resources |
| Scaling law |
In the context of machine learning, an empirical relationship between a model's size metrics-such as number of weights, number of tokens in the training dataset, and the floating-point operations used for training-and the model's performance as measured by a loss function. Scaling laws are widely used to justify investments in larger and larger models. |
Jared Kaplan et al., "Scaling Laws for Neural Language Models," arXiv preprint, submitted January 23, 2020, https://arxiv.org/abs/2001.08361.
Jordan Hoffmann et al., "Training Compute-Optimal Large Language Models," arXiv preprint, submitted March 29, 2022,
https://doi.org/10.48550/arXiv.2203.15556. |
Hardware and Computational Resources |
| Self-attention |
In the context of artificial intelligence and transformer models more specifically, a type of attention mechanism that maps weights onto terms according to their different positions within a single sequence in order to compute a representation of the associations between terms in a sequence. |
Ashish Vaswani at al., "Attention Is All You Need," Advances in Neural Information Processing Systems 30 (2017): p. 2. |
Neural Network Components and Functions |
| Semantic search |
"Semantic search uses vector search to deliver and rank content based on context relevance and intent relevance. Vector search encodes details of searchable information into vectors, and then compares vectors to determine which are most similar. A vector-search-enabled semantic search produces results by working at both ends of the query pipeline simultaneously: When a query is launched, the search engine transforms the query into embeddings, which are numerical representations of data and related contexts. They are stored in vectors. The kNN algorithm, or k-nearest neighbor algorithm, then matches vectors of existing documents (a semantic search concerns text) to the query vectors. The semantic search then generates results and ranks them based on conceptual relevance." |
"What Is Semantic Search?" Elastic, https://www.elastic.co/what-is/semantic-search. |
Applications and Use Cases |
| Start-of-sequence token |
In the context of generative artificial intelligence, a string of characters a model generates that lets the model know to start generating text. See also "end-of-sequence token." |
Ashton Zhang et al., "Sequence to Sequence Learning," in Dive Into Deep Learning, https://classic.d2l.ai/chapter_recurrent-modern/seq2seq.html. |
Natural Language Processing |
State-space model
|
A class of probabilistic graphical models that describe the probabilistic dependence between a latent state variable and observed measurements. State-space models were popularized in the field of control engineering in the second half of the 20th century and have since found use in machine learning. See "MAMBA." |
Albert Gu, Karan Goel, Khaled Saab, and Chris Re, "Structured State Spaces: Combining Continuous-Time, Recurrent, and Convolutional Models," Hazy Research, updated January 14, 2022, https://hazyresearch.stanford.edu/blog/2022-01-14-s4-3. |
AI Model Types and Architectures |
| Stop sequence |
A synonym for "end-of-sequence token." |
Ashton Zhang et al., "Sequence to Sequence Learning," in Dive Into Deep Learning, https://classic.d2l.ai/chapter_recurrent-modern/seq2seq.html. |
Natural Language Processing |
| Summarization |
In the context of artificial intelligence, the ability of an artificial intelligence model to perform the task of reducing a text into a shorter text while keeping all (or most) of the important aspects referenced in the original text, whether through abstractive techniques or extractive techniques. |
"Transformers: What Can They Do?" Hugging Face, https://huggingface.co/learn/nlp-course/chapter1/3?fw=pt. |
Applications and Use Cases |
| Supervised fine-tuning (SFT) |
A fine-tuning method that presents example prompts and completions to a model and adjusts the model's weights such that it is more likely to imitate the demonstrated patterns. Supervised fine-tuning datasets commonly number in the tens of thousands of examples and may be produced either manually or by another model. |
Chip Huyen, "RLHF: Reinforcement Learning from Human Feedback," updated May 2, 2023, https://huyenchip.com/2023/05/02/rlhf.html. |
Learning Methods and Techniques |
| Synthetic content |
In the context of artificial intelligence, "information, such as images, videos, audio clips, and text, that has been significantly modified or generated by algorithms, including by AI." |
Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence, Section 3(ee), October 20, 2023, https://www.whitehouse.gov/briefing-room/presidential-actions/2023/10/30/executive-order-on-the-safe-secure-and-trustworthy-development-and-use-of-artificial-intelligence/. |
Applications and Use Cases |
| Tagging |
A synonym for "annotation." |
"Data Annotation and Labeling: Everything You Need to Know," Amantya Technologies, https://www.amantyatech.com/data-annotation-and-labeling-everything-you-need-to-know. |
Applications and Use Cases |
| Temperature |
In the context of generative artificial intelligence, a specific user-assigned model parameter that determines the randomness the model will employ when generating outputs. |
Harry Guinness, "How to Use the OpenAI Playground to Tinker with GPT-3 and GPT-4," Zapier, updated July 16, 2023, https://zapier.com/blog/openai-playground/. |
Sampling Methods |
| Tensor processing unit (TPU) |
An "application-specific integrated circuit" processor Google designed to serve as an alternative to a "graphics processing unit" for training and inference of large artificial intelligence models. |
"Accelerate AI development with Google Cloud TPUs," Google Cloud, https://cloud.google.com/tpu. |
Hardware and Computational Resources |
| Testbed |
In the context of artificial intelligence, "a facility or mechanism equipped for conducting rigorous, transparent, and replicable testing of tools and technologies, including AI and PETs, to help evaluate the functionality, usability, and performance of those tools or technologies." |
Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence, Section 3(ff), October 20, 2023, https://www.whitehouse.gov/briefing-room/presidential-actions/2023/10/30/executive-order-on-the-safe-secure-and-trustworthy-development-and-use-of-artificial-intelligence/. |
Performance and Evaluation |
| The Pile |
A common training data resource for large language models. It is "an 825 GiB English text corpus targeted at training large-scale language models. The Pile is constructed from 22 diverse high-quality subsets-both existing and newly constructed-many of which derive from academic or professional sources." |
Leo Gao et al., "The Pile: An 800GB Dataset of Diverse Text for Language Modeling," arXiv preprint, submitted December 31, 2020, https://arxiv.org/abs/2101.00027. |
Datasets |
| Token |
In the context of artificial intelligence and natural language processing more specifically, "a sequence of characters (not necessarily a whole natural word) grouped together as a useful semantic unit for processing." A rule-of-thumb is that 100 tokens is equal to about 75 words of English text. Tokens can also be created for other data modalities, such as images. In the case of images, a token represents a patch of pixels. |
Christopher Manning, Prabakhar Raghavan, Hinrich Schutze. 2008. Introduction to Information Retrieval. Available at: https://nlp.stanford.edu/IR-book/html/htmledition/tokenization-1.html |
Natural Language Processing |
| Tokenization |
In the context of artificial intelligence, the process of creating tokens or discrete chunks of data from natural language text sequences. |
Amal Menzli, "Tokenization in NLP: Types, Challenges, Examples, Tools," MLOps Blog, August 11, 2023, https://neptune.ai/blog/tokenization-in-nlp. |
Natural Language Processing |
| Top-k sampling |
A configuration parameter for large language models that instructs a model to choose the top k most probable next tokens when generating output and sample among those only. It can reduce the model's likelihood of generating less human-sounding text. |
Fabio Chiusano, "Two Minutes NLP - Most Used Decoding Methods for Language Models," NLPlanet, updated January 28, 2022, https://medium.com/nlplanet/two-minutes-nlp-most-used-decoding-methods-for-language-models-9d44b2375612. |
Sampling Methods |
| Top-p sampling |
A configuration parameter for large language models that instructs a model to sample only next tokens whose cumulative probability exceeds a threshold p. It can reduce the model's likelihood of generating less human-sounding text. Unlike in top-k sampling, the number of tokens in the sampling set can vary dynamically as the sequence is generated. |
Fabio Chiusano, "Two Minutes NLP - Most Used Decoding Methods for Language Models," NLPlanet, updated January 28, 2022, https://medium.com/nlplanet/two-minutes-nlp-most-used-decoding-methods-for-language-models-9d44b2375612. |
Sampling Methods |
| Toxicity |
In the context of generative artificial intelligence, any one or all of multiple measures of an artificial intelligence model's capability to identify data as rude, profane, hateful, pornographic, or disrespectful in nature, or to respond appropriately to remove or limit access to these data. |
Sohl Dickstein, "BIG-Bench Keywords," GitHub, updated June 4, 2022, https://github.com/google/BIG-bench/blob/main/keywords.md. |
AI Ethics and Governance |
| Traceable |
A principle for ensuring the trustworthiness of artificial intelligence. Traceable artificial intelligence systems are developed and deployed such that relevant personnel possess an appropriate understanding of the technology, development processes, and operational methods applicable to the artificial intelligence's capabilities, including with transparent and auditable methodologies, data sources, and design procedures and documentation. |
Department of Defense, "Responsible Artificial Intelligence Strategy and Implementation Pathway" (Washington, DC: USD(AT&L), June 2022), https://media.defense.gov/2022/Jun/22/2003022604/-1/-1/0/Department-of-Defense-Responsible-Artificial-Intelligence-Strategy-and-Implementation-Pathway.PDF. |
AI Ethics and Governance |
| Transfer learning |
In the context of artificial intelligence, the multiple acts required to initialize a new model with another model's weights so that a model capable of performing one task becomes capable of performing another task. |
"How Do Transformers Work?" Hugging Face, https://huggingface.co/learn/nlp-course/chapter1/4?fw=pt. |
Learning Methods and Techniques |
| Transformer model |
In the context of artificial intelligence and neural networks more specifically, a type of neural network model that learns context and meaning by mapping relationships in sequential data using attention mechanisms, encoders, and decoders. Leading large language models as of December 2023 are all transformer models. |
Rick Merritt, "What Is a Transformer Model?" NVIDIA, updated March 25, 2022, https://blogs.nvidia.com/blog/2022/03/25/what-is-a-transformer-model. |
AI Model Types and Architectures |
| Translation |
The transformation of text from one natural language to another in a way that preserves context, sentiment, and meaning. |
"What Is Machine Translation?" Amazon Web Services, https://aws.amazon.com/what-is/machine-translation/. |
Applications and Use Cases |
| Tree-of-thoughts (ToT) prompting |
A framework that generalizes over chain-of-thought prompting and encourages exploration via thoughts that serve as intermediate steps for general problem-solving with language models. In ToT prompting, thoughts represent coherent language sequences that serve as intermediate steps toward solving a problem. This approach enables a language model to self-evaluate the progress intermediate thoughts make toward solving a problem through a deliberate reasoning process. The language model's ability to generate and evaluate thoughts is then combined with search algorithms to enable systematic exploration of these intermediate steps. |
"Tree of Thoughts," Prompt Engineering Guide, DAIR.AI, https://www.promptingguide.ai/techniques/tot. |
Prompting Technique |
| Trustworthy |
Characteristics of trustworthy AI systems include: valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed. Addressing AI trustworthiness characteristics individually will not ensure AI system trustworthiness; tradeoffs are usually involved, rarely do all characteristics apply in every setting, and some will be more or less important in any given situation. Ultimately, trustworthiness is a social concept that ranges across a spectrum and is only as strong as its weakest
characteristics. |
National Institute of Standards and Technology, "Artificial Intelligence Risk Management Framework (AI RMF 1.0)," January, 2023, https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf. |
AI Ethics and Governance |
| Truthfulness |
In the context of artificial intelligence, "truthfulness means that the AI system's statements truthfully describe the world." |
Owain Evans et al., "Truthful AI: Developing and Governing AI That Does Not Lie," arXiv preprint, submitted October 13, 2021, https://arxiv.org/pdf/2110.06674.pdf. |
AI Ethics and Governance |
| Value function |
In the context of artificial intelligence and reinforcement learning more specifically, the objective a reinforcement learning agent is to achieve. The value function and the "reward function" differ in terms of the length of their horizon for the agent: value is long-term gain and reward is short-term gain. |
"Elements of Reinforcement Learning," in Richard S. Sutton and Andrew G. Barto, Reinforcement Learning: An Introduction (Cambridge, MA: MIT Press, 1998), http://incompleteideas.net/book/ebook/node9.html. |
Learning Methods and Techniques |
| Variational autoencoder (VAE) |
A type of model architecture composed of encoder and decoder models wherein the decoder model must minimize the reconstruction error between algorithmically transformed data and the input data. VAEs encode input data as a distribution over a latent space rather than as a single point, incorporate a reconstruction term on the output later into the loss function, and incorporate a regularization term on the latent later that regularizes the organization of the latent space. VAEs are a component in popular diffusion models. |
Joseph Rocca, "Understanding Variational Autoencoders (VAEs): Building, Step by Step, the Reasoning That Leads to VAEs," Towards Data Science, September 23, 2019, https://towardsdatascience.com/understanding-variational-autoencoders-vaes-f70510919f73. |
AI Model Types and Architectures |
| Vector database |
A database designed to store embeddings and perform operations with them. Vector databases are commonly used in retrieval augmented generation systems for storing embeddings of relevant documents that are too large to fit in a large language model's context window. See "embedding." |
Roie Schwaber-Cohen, "What Is a Vector Database and How Does It Work? Use Cases + Examples," Pinecone, https://www.pinecone.io/learn/vector-database/. |
Data Handling and Processing |
| Visual question answer (VQA) |
The ability of an artificial intelligence model to answer questions about an image in natural language. |
"Computer Vision," Papers with Code, https://paperswithcode.com/task/visual-question-answering. |
Applications and Use Cases |
| Watermarking |
In the context of artificial intelligence, a technique that involves embedding a hidden signal, such as a pixel, to identify that content was generated by artificial intelligence when read by a corresponding piece of software. Such markings should be reliable, accessible, and difficult to remove. |
Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence, Section 3(gg), October 20, 2023, https://www.whitehouse.gov/briefing-room/presidential-actions/2023/10/30/executive-order-on-the-safe-secure-and-trustworthy-development-and-use-of-artificial-intelligence/.
Apostol Vassilev, Alina Oprea, Alie Fordyce, and Hyrum Anderson, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations, NIST Trustworthy and Responsible AI, NIST AI 100-2e2023 (Gaithersburg, MD: National Institute of Standards and Technology, January 2024), https://doi.org/10.6028/NIST.AI.100-2e2023. |
AI Ethics and Governance |
| World model |
In the context of generative artificial intelligence, the ability of a large language model to build or maintain a representation of past and future states of the relationship between acts, objects, and time. |
David Ha and Jurgen Schmidhuber, "World Models: Can Agents Learn Inside of Their Own Dreams?" March 27, 2018, NIPS 2018, https://worldmodels.github.io/. |
Other |
| Zero-shot prompting |
A special case of N-shot prompting wherein the user prompts a large language model for output without providing any examples. Also spelled "0-shot prompting." |
Sunil Ramlochan, "Master Prompting Concepts: Zero-Shot and Few-Shot Prompting," Prompt Engineering Institute, April 25, 2023, https://promptengineering.org/master-prompting-concepts-zero-shot-and-few-shot-prompting/. |
Prompting Technique |