Resources

AI Glossary

Plain-language definitions of the AI, automation, and agentic development terms you actually need to know.

399 terms

A

A-Star Search: A graph exploration and route-finding algorithm widely applied in various areas of computer science because of its thoroughness, optimality, and efficient performance.
Abductive Logic Programming (ALP): A broad knowledge-representation paradigm that facilitates problem-solving in a declarative manner using abductive inference. It eases conventional logic programming by permitting certain predicates to remain partially specified, designating them as abducible predicates.
Abductive Reasoning: A type of logical reasoning that begins with an observation or a series of observations and then attempts to determine the most straightforward and probable explanation. Unlike deductive logic, this method produces a likely conclusion without definitively confirming it. Also known as abductive inference or retroduction.
Ablation: The elimination of a specific element within an AI system. An ablation study seeks to assess the impact of that element by extracting it and subsequently evaluating the system's performance to determine its contribution.
Abstract Data Type: A formal mathematical framework for data types, wherein a data type is characterized by its functional behavior from the perspective of a user. This is defined in terms of permissible values, applicable operations on data of this type, and the expected behavior of those operations.
Abstraction: The procedure of eliminating physical, spatial, or temporal specifics or attributes in the examination of objects or systems to focus more precisely on other aspects of significance.
Accelerating Change: An observed acceleration in the pace of technological advancement, potentially indicating more rapid and significant transformations in the future, which may or may not be accompanied by equally substantial social and cultural shifts.
Accuracy: A measurement metric in binary classification determined using the formula: (True Positives + True Negatives) / (True Positives + True Negatives + False Positives + False Negatives).
Actionable Intelligence: Data that can be utilized to aid in decision-making.
Action Language: A formal notation for defining state transition systems, frequently utilized to construct precise models of how actions influence the surrounding environment. Action languages are widely applied in artificial intelligence and robotics, where they depict the impact of actions on system states over time and can be employed for automated decision-making and planning.
Action Model Learning: A branch of machine learning focused on developing and refining a software agent's understanding of the outcomes and prerequisites of actions it can perform within its surroundings. This knowledge is typically encoded in a logic-based action description language and serves as input for automated planning systems.
Action Selection: A method of defining the fundamental challenge faced by intelligent systems: determining the next course of action. In artificial intelligence and computational cognitive science, the action selection problem is commonly linked to intelligent agents and animats — artificial entities that demonstrate intricate behaviors within an agentic environment.
Activation Function: In artificial neural networks, a node's activation function determines its output based on a given input or collection of inputs.
Adaptive Algorithm: An algorithm that adapts its behavior during execution based on a predefined reward system or evaluation criterion.
Adaptive Learning: An approach that leverages data-informed instruction to modify and customize educational experiences to suit the unique requirements of each learner. Adaptive learning platforms can monitor a student's advancement, involvement, and achievement, utilizing this data to deliver individualized learning experiences.
Adaptive Neuro Fuzzy Inference System (ANFIS): A type of artificial neural network derived from the Takagi–Sugeno fuzzy inference system, developed in the early 1990s. It merges neural networks with fuzzy logic concepts, harnessing the advantages of both within a unified framework. Its inference mechanism consists of fuzzy IF-THEN rules with the ability to learn and approximate nonlinear functions. Consequently, ANFIS is regarded as a universal estimator. Also referred to as an adaptive network-based fuzzy inference system.
Admissible Heuristic: A heuristic function is considered admissible if it never predicts a cost exceeding the actual minimum cost required to reach the goal. In other words, the estimated cost to reach the objective is always less than or equal to the least possible cost from the current position along the path.
Agent Architecture: A framework for software agents and intelligent control systems, illustrating the organization of components. The structures utilized by intelligent agents are known as cognitive architectures.
AI Accelerator: A category of microprocessors or computing systems engineered as specialized hardware to improve the performance of AI tasks, particularly artificial neural networks, computer vision, and machine learning.
AI-Complete: In artificial intelligence, the most challenging problems are informally termed AI-complete or AI-hard, indicating that their complexity is on par with solving the fundamental AI challenge — creating machines with human-level intelligence (strong AI). Labeling a problem as AI-complete suggests that it cannot be resolved using a straightforward, specialized algorithm.
AI Ethics: The field encompassing concerns that AI stakeholders, including engineers and policymakers, must address to guarantee the responsible development and deployment of AI technology. This concept involves designing and applying systems that promote a safe, reliable, impartial, and environmentally sustainable approach to AI.
Algorithm: A set of instructions provided to an AI system to carry out a task or resolve a problem. Common types of computer algorithms include classification, regression, and clustering.
Algorithmic Efficiency: A characteristic of an algorithm that pertains to the quantity of computational resources it consumes. An algorithm must undergo analysis to assess its resource utilization, and its effectiveness can be evaluated based on the consumption of various resources.
Algorithmic Probability: In algorithmic information theory, also referred to as Solomonoff probability, a mathematical approach for allocating a prior likelihood to a specific observation. This concept was devised by Ray Solomonoff during the 1960s.
AlphaGo: A software application designed to play the board game Go, created by Google DeepMind. AlphaGo has multiple iterations including AlphaGo Zero, AlphaGo Master, and AlphaGo Lee. In October 2015, AlphaGo became the first AI-powered Go program to defeat a human professional Go player without handicaps on a standard 19×19 board.
Ambient Intelligence (AmI): Digital surroundings that are perceptive and reactive to human presence.
Analysis of Algorithms: The assessment of an algorithm's computational difficulty, specifically the quantity of time, memory, and other resources required for execution. This typically entails identifying a function that correlates the size of an algorithm's input to the number of operations it performs (temporal complexity) or the amount of storage it consumes (spatial complexity).
Analytics: The identification, analysis, and conveyance of significant trends within data.
Anaphora: In linguistics, a pronoun used to refer back to a previously mentioned noun. For instance, in the sentence 'While Paul didn't like the appetizers, he enjoyed the entrée,' the word 'he' serves as the anaphora.
Annotation: The method of labeling linguistic data by recognizing and marking grammatical, semantic, or phonetic components within the language data.
Answer Set Programming (ASP): A type of declarative programming designed for tackling complex, mainly NP-hard, search challenges. It is founded on the stable model (answer set) semantics of logic programming. In ASP, search problems are transformed into the task of computing stable models, with answer set solvers used to execute the search process.
Ant Colony Optimization (ACO): A stochastic method for resolving computational challenges that can be transformed into the task of identifying optimal routes within graphs.
Anytime Algorithm: An algorithm capable of producing a legitimate solution to a problem, even if halted before reaching completion.
Application Programming Interface (API): A collection of protocols that define how two software applications communicate with one another. APIs are typically written in programming languages like C++ or JavaScript.
Approximate String Matching: The method of identifying strings that closely resemble a specified pattern rather than matching it precisely. The challenge is generally categorized into two sub-problems: locating near-matching substrings within a given string and retrieving dictionary entries that approximately correspond to the pattern.
Approximation Error: The deviation between a precise value and an estimated approximation of it.
Argumentation Framework: A method for handling disputed information and deriving conclusions from it. Within an abstract argumentation framework, foundational data consists of a set of abstract arguments depicting facts or assertions, with disputes between arguments represented through a binary relationship. In practical terms, it is illustrated as a directed graph where nodes signify arguments and arrows denote the attack relation.
Artificial General Intelligence (AGI): A form of artificial intelligence that equals or exceeds human intellectual abilities within a broad spectrum of mental tasks.
Artificial Immune System (AIS): A category of computationally intelligent, rule-based machine learning frameworks influenced by the mechanisms and functions of the vertebrate immune system. These algorithms are generally designed to replicate the immune system's adaptive learning and memory capabilities for problem-solving applications.
Artificial Intelligence (AI): The emulation of human cognitive processes by machines or computer systems. AI can replicate human abilities such as communication, learning, and decision-making.
Artificial Intelligence Markup Language: An XML-based language variant designed for developing intelligent software entities that process natural language.
Artificial Narrow Intelligence (ANI): AI that can address specific issues rather than general tasks. For instance, a smartphone can utilize facial recognition to detect images of a person, but that same system is unable to recognize sounds.
Artificial Neural Network (ANN): Often called a neural network, this system comprises a network of interconnected nodes or units that roughly emulate the processing capabilities of the human brain.
Association for the Advancement of Artificial Intelligence (AAAI): A global, nonprofit scientific organization dedicated to advancing research in and the ethical application of AI, expanding public awareness of AI, and refining the education and training of AI professionals.
Asymptotic Computational Complexity: In computational complexity theory, the application of asymptotic analysis to estimate the computational demands of algorithms and computational challenges, typically linked to the use of Big O notation.
Attention Mechanism: A machine learning-based system that emulates human cognitive focus. It assigns soft weights to each word (its embedding) within a contextual window. This process can be performed concurrently (as in transformers) or sequentially (as in recursive neural networks). Unlike hard weights which remain fixed, soft weights can dynamically adjust during each execution.
Attributional Calculus: A reasoning and representation framework formulated by Ryszard S. Michalski, integrating aspects of predicate logic, propositional calculus, and multi-valued logic. It serves as a structured language for natural induction, an inductive learning approach that yields results in a format intuitive to humans.
Augmented Reality (AR): An immersive interaction with a real-world setting where elements from the physical environment are enhanced by digitally generated perceptual information, often spanning multiple sensory channels such as visual, auditory, haptic, somatosensory, and olfactory stimuli.
Auto-Classification: The use of a machine learning system, natural language processing (NLP), and other AI methods to automate text classification with greater speed, efficiency, and accuracy.
Auto-Complete: A search feature that proposes potential queries based on the text being entered to formulate a search request.
Autoencoder: A category of artificial neural networks designed to acquire optimized representations of unlabelled data through unsupervised learning. A widely used implementation of this approach is the variational autoencoder (VAE).
Automata Theory: The exploration of theoretical machines and automata, along with the computational challenges that can be addressed using them. This field is a branch of theoretical computer science and discrete mathematics.
Automated Machine Learning (AutoML): A branch of machine learning focused on the automatic optimization of an ML system to increase its effectiveness.
Automated Planning and Scheduling: A subdivision of artificial intelligence focused on devising strategies or action sequences, usually intended for implementation by intelligent agents, self-governing robots, and autonomous vehicles. Unlike conventional control and classification challenges, solutions must be identified and refined within a multidimensional space. Planning is also linked to decision theory.
Automated Reasoning: A domain within computer science and mathematical logic focused on analyzing various facets of reasoning. Research in automated reasoning facilitates the development of software that helps computers infer conclusions fully or almost entirely autonomously.
Autonomic Computing (AC): The self-regulating attributes of decentralized computing resources that adjust to unforeseen variations while concealing underlying complexity from administrators and end-users. Introduced by IBM in 2001, this initiative was intended to create computing systems with autonomous management capabilities.
Autonomous Car: A transport system capable of perceiving its surroundings and navigating with minimal or no human intervention.
Autonomous Robot: A machine that executes actions or functions with a significant level of independence. Autonomous robotics is generally regarded as a branch of AI, robotics, and information engineering.

B

Backpropagation: A technique utilized in artificial neural networks to determine a gradient essential for computing the weights applied within the network. Backpropagation is short for backward propagation of errors — the error is measured at the output and propagated in reverse through the network's layers. It is frequently employed to train deep neural networks containing multiple hidden layers.
Backpropagation Through Structure (BPTS): A gradient-based method for optimizing recurrent neural networks, introduced in a 1996 research paper authored by Christoph Goller and Andreas Küchler.
Backpropagation Through Time (BPTT): A gradient-based technique for training recurrent neural networks by unrolling the network through time and applying backpropagation across the temporal sequence of operations.
Backward Chaining: A reasoning technique commonly referred to as operating in reverse from the desired outcome. It is employed in automated proof systems, deduction engines, verification assistants, and various artificial intelligence applications.
Bag-of-Words Model: A streamlined representation utilized in natural language processing and information retrieval. Within this framework, a text is depicted as a collection of its words, ignoring syntax and word sequence while preserving word frequency. It is frequently employed in document categorization techniques, where the presence and frequency of each word serve as attributes for training a classifier.
Bag-of-Words Model in Computer Vision: In computer vision, the bag-of-words (BoW) approach can be utilized for image categorization by interpreting image attributes as words. In text classification, a bag of words represents a sparse vector containing word occurrence frequencies. Similarly, in computer vision, a bag of visual words consists of a vector quantifying the frequency of a predefined set of local image descriptors.
Batch Normalization: A method for enhancing the efficiency and robustness of artificial neural networks by supplying any layer with inputs that have zero mean and unit variance. It functions by standardizing the input layer through modification and rescaling of activations. First presented in a 2015 research paper.
Bayesian Programming: A framework and approach for establishing a method to define probabilistic models and address problems when incomplete information is present.
Bees Algorithm: A population-based optimization technique introduced by Pham, Ghanbarzadeh, and collaborators in 2005. It emulates the food-seeking behavior of honeybee swarms. The algorithm integrates localized exploration with broad search strategies and is applicable to both combinatorial and continuous optimization problems.
Behavior Informatics (BI): The analysis of behavioral data to extract intelligence and insights regarding actions and tendencies.
Behavior Tree (BT): A mathematical framework for executing plans, widely applied in computer science, robotics, control systems, and video games. It models transitions between a finite collection of tasks in a structured manner, enabling the construction of highly intricate behaviors from simpler tasks. Behavior Trees share similarities with hierarchical state machines, with the key distinction that their fundamental unit is a task rather than a state.
Belief–Desire–Intention Software Model (BDI): A computational framework designed for developing intelligent agents. It incorporates an agent's beliefs, desires, and intentions to address specific challenges in agent-based programming. At its core, it establishes a mechanism for distinguishing between the selection of a plan and the execution of currently active plans, allowing BDI agents to balance time allocated to deliberating over plans and carrying them out.
Bias-Variance Tradeoff: In statistics and machine learning, the characteristic of a collection of predictive models in which models exhibiting lower bias in parameter estimation tend to have greater variability in parameter estimates across different samples, and conversely, models with higher bias demonstrate reduced variance.
Bidirectional Coder Representation From Transformers (BERT): A large-scale pre-trained model initially trained on vast amounts of unlabelled data. It is then adapted to a specific NLP task by being provided with a smaller, task-focused dataset for fine-tuning the final model.
Big Data: Extensive data sets that can be analyzed to uncover patterns and trends that inform business decisions. Organizations can now accumulate vast and intricate data using various collection tools and systems, gathering it rapidly and storing it in multiple formats.
Big O Notation: A symbolic representation that characterizes the asymptotic behavior of a function as its input approaches a specific value or extends to infinity. It belongs to a group of notations introduced by Paul Bachmann, Edmund Landau, and others, collectively referred to as Bachmann–Landau notation or asymptotic notation.
Binary Tree: A hierarchical data structure where each element (a node) has at most two subordinate nodes, termed the left child and the right child. A recursive characterization using set theory defines a non-empty binary tree as an ordered triple (L, S, R), where L and R represent either binary trees or an empty set, and S is a singleton set.
Blackboard System: An artificial intelligence methodology rooted in the blackboard architectural paradigm, wherein a shared information repository (the blackboard) is progressively modified by a diverse set of expert knowledge modules. The process begins with a problem definition and concludes with a resolved outcome.
Black Boxes: Things we do not comprehend because their inner workings are hidden from view. Many machine learning models are considered black boxes, meaning we lack insight into how they utilize different aspects of the data when making decisions. Two main approaches to uncovering the inner workings of AI models are interpretable machine learning and explainable machine learning.
Boltzmann Machine: A category of probabilistic recurrent neural network and Markov random field, regarded as the stochastic, generative equivalent of Hopfield networks.
Boolean Satisfiability Problem: The challenge of determining whether there exists an assignment of values that satisfies a specified Boolean expression — whether the variables within a given Boolean formula can be consistently substituted with TRUE or FALSE such that the expression evaluates to TRUE. If such an assignment exists, the formula is termed satisfiable; otherwise, it is deemed unsatisfiable.
Boosting: A machine learning ensemble optimization technique primarily aimed at minimizing bias rather than variance by training models in a sequential manner, with each successive model adjusting for the errors made by its predecessor.
Bootstrap Aggregating: A machine learning ensemble optimization method primarily designed to decrease variance, rather than bias, by training multiple models separately and combining their predictions through averaging.
Brain Technology: A technology that leverages the most recent advancements in neuroscience. The term was originally coined by the Artificial Intelligence Laboratory in Zurich, Switzerland, within the scope of the ROBOY project. It can be utilized in robotics, knowledge management systems, and various other applications with self-learning abilities.
Branching Factor: In computing, hierarchical data structures, and game theory, the quantity of descendant nodes at each parent node, known as the outdegree. If this number varies by node, an average branching factor can be determined.
Brute-Force Search: A highly versatile problem-solving method and algorithmic framework that involves methodically listing all potential solution candidates and verifying whether each one meets the conditions specified by the problem.

C

Capsule Neural Network (CapsNet): A machine learning framework that represents a category of ANNs designed to more effectively capture hierarchical relationships. This methodology seeks to emulate the structural organization of biological neural systems more accurately.
Case-Based Reasoning (CBR): The method of resolving novel problems by leveraging solutions from previously encountered, analogous problems.
Cataphora: In linguistics, a reference that appears before the noun it refers to. For example, in the sentence 'Though he enjoyed the entrée, Paul didn't like the appetizers,' the word 'he' functions as a cataphora.
Categorization: A natural language processing capability that assigns a specific category to a document.
Category: A designation given to a document to characterize the information it contains.
Category Trees: Allows viewing all rule-based classifications within a collection. Used to create, remove, and modify the rules that link documents to categories. Also known as a taxonomy, structured in a hierarchical format.
Chat-Based Generative Pre-Trained Transformer (ChatGPT) Models: A system constructed with a neural network transformer-style AI model that excels in natural language processing tasks. The model can produce replies to inquiries (Generative), was pre-trained on a substantial amount of textual data from the internet (Pre-trained), and can analyze sentences in a way that differs from other model types (Transformer).
Chatbot: A software program created to simulate human dialogue using text or voice inputs.
Classification: Methods that allocate a set of predefined labels to unstructured text, allowing for the organization, structuring, and classification of various types of text — from documents and medical records to emails and files.
Cloud Robotics: A branch of robotics that seeks to integrate cloud-based technologies — including cloud computing, cloud storage, and various Internet-based solutions — to harness the advantages of converged infrastructure and shared services for robotic applications. Robots can utilize the vast computational power, storage capacity, and communication capabilities of modern cloud data centers, enabling lightweight, cost-efficient, and more intelligent robots.
Cluster Analysis: The process of organizing a collection of items into subsets (clusters) such that elements within the same cluster exhibit greater similarity compared to those in different clusters. A fundamental operation in exploratory data mining with applications in machine learning, pattern recognition, image processing, information retrieval, bioinformatics, data compression, and computer graphics.
COBWEB: A progressive framework for hierarchical conceptual grouping, devised by Professor Douglas H. Fisher. This method incrementally structures observations into a taxonomy tree, where each node signifies a category characterized by a probabilistic representation encapsulating the attribute-value distributions of entities classified under it.
Cognitive Architecture: A theoretical framework regarding the immutable structures that constitute a mind — whether in biological or synthetic systems — and the manner in which they interact alongside the knowledge and competencies embedded within the framework to produce intelligent conduct in intricate settings.
Cognitive Computing: Essentially synonymous with artificial intelligence. A computational model designed to replicate human cognitive processes, including pattern recognition and learning.
Cognitive Map: Also referred to as a mental palace, an internal representation that helps an individual absorb, encode, retain, retrieve, and interpret information regarding the spatial relationships and characteristics of elements within their surroundings.
Cognitive Science: The cross-disciplinary scientific investigation of cognition and its mechanisms.
Combinatorial Optimization: In operations research, applied mathematics, and theoretical computer science, a field focused on identifying the best possible element from a limited collection of choices.
Committee Machine: A category of artificial neural network that employs a divide-and-conquer approach, where the outputs of multiple neural networks are integrated into a unified response. The aggregated output is intended to surpass the performance of its individual expert components.
Commonsense Knowledge: In artificial intelligence studies, universally known facts about the everyday world that all people are presumed to understand. The earliest AI system designed to handle commonsense knowledge was Advice Taker, developed by John McCarthy in 1959.
Commonsense Reasoning: A subdivision of AI focused on replicating the human capacity to infer assumptions regarding the nature and characteristics of commonplace situations encountered daily.
Completions: The result produced in response to a generative input.
Composite AI: The integrated use of various AI methods to increase learning efficiency, expand the scope of knowledge representations, and ultimately address a broader array of business challenges more effectively.
Computational Chemistry: A field of chemistry that employs computational modeling to aid in resolving chemical challenges.
Computational Complexity Theory: Centers on categorizing computational challenges based on their intrinsic complexity and establishing relationships among these categories. A computational challenge refers to a task executed by a computer that can be resolved through the systematic application of mathematical procedures such as an algorithm.
Computational Creativity: A cross-disciplinary pursuit encompassing the domains of AI, cognitive psychology, philosophy, and the creative arts.
Computational Cybernetics: The fusion of cybernetics with computational intelligence methodologies.
Computational Humor: A subfield of computational linguistics and artificial intelligence that leverages computers for the study of humor.
Computational Intelligence (CI): Typically denotes a computer's capability to acquire knowledge of a particular task through data analysis or empirical observation.
Computational Learning Theory: A branch of AI focused on examining the development and evaluation of machine learning algorithms.
Computational Linguistics: A multidisciplinary domain focused on the algorithmic representation and processing of natural language.
Computational Mathematics: The mathematical investigation in scientific domains where computation serves a fundamental role.
Computational Neuroscience: A field of neuroscience that utilizes mathematical frameworks, theoretical examination, and abstractions of the brain to comprehend the fundamental principles guiding the formation, organization, functionality, and cognitive capabilities of the nervous system.
Computational Number Theory: The exploration of computational methods for executing arithmetic operations within number theory.
Computational Problem: In theoretical computing, a mathematical construct that encapsulates a set of queries that computers may be capable of resolving.
Computational Semantics: The exploration of methods to automate the creation and interpretation of meaning representations for natural language expressions.
Computational Statistics: The juncture between computational science and statistical analysis.
Computer-Automated Design (CAutoD): Expanding upon Computer-Aided Design (CAD), automated design encompasses a wider array of applications including automotive engineering, civil engineering, composite material development, control systems engineering, and structural optimization. CAutoD evolves from conventional CAD simulation through biologically inspired machine learning, incorporating heuristic search strategies such as evolutionary computation and swarm intelligence algorithms.
Computer Science: The principles, practical investigation, and applied engineering that underpin the conception and utilization of computing devices. It involves developing algorithms that manipulate, retain, and transmit digital data.
Computer Vision: A multidisciplinary domain within science and technology that concentrates on helping computers interpret and extract insights from images and videos, facilitating the automation of tasks traditionally carried out by the human visual system.
Concept Drift: In prognostic analytics and machine learning, the phenomenon where the statistical characteristics of the target variable that the model aims to forecast evolve unpredictably, causing accuracy of predictions to deteriorate progressively over time.
Connectionism: A methodology within cognitive science that seeks to elucidate mental phenomena through the utilization of artificial neural networks.
Consistent Heuristic: In the exploration of pathfinding challenges within artificial intelligence, a heuristic function is considered consistent or monotonic if its estimation never exceeds the projected distance from any adjacent vertex to the destination, combined with the cost required to reach that neighbor.
Constrained Conditional Model (CCM): A machine learning and reasoning framework that improves the training of conditional models by incorporating declarative constraints.
Constraint Logic Programming: A variant of constraint programming where logic programming is expanded to incorporate principles from constraint satisfaction. A constraint logic program consists of a logic-based framework that integrates constraints within clause bodies.
Constraint Programming: A programming model in which relationships among variables are expressed as constraints. Unlike imperative programming languages, constraints do not define a sequence of operations to perform but instead describe the characteristics that a valid solution must satisfy.
Constructed Language: A linguistic system whose sound patterns, syntax, and lexicon are intentionally created rather than evolving organically. Also known as artificial, designed, or fabricated languages.
Content: Distinct units of information — namely, documents — that can be aggregated to create training datasets or produced by generative AI.
Content Enrichment: The method of utilizing sophisticated approaches like machine learning, AI, and natural language processing to automatically derive valuable insights from text-based documents.
Controlled Vocabulary: A carefully selected set of terms and expressions pertinent to a particular application or industry. These elements may include additional attributes that define their linguistic behavior and the meanings they convey in relation to topics and other contexts.
Control Theory: In control systems engineering, a specialized branch of mathematics that focuses on regulating continuously functioning dynamic systems within designed processes and machinery, with the goal of formulating a control framework that minimizes delay and overshoot while maintaining stability.
Conversational AI: Platforms utilized by developers to create interactive user experiences, chatbots, and virtual assistants for many applications, supporting integration with communication channels such as messaging apps, social networks, SMS, and websites.
Convolutional Neural Networks (CNN): A category of deep learning neural networks comprising one or multiple layers, designed for image analysis and recognition.
Co-Occurrence: The appearance of distinct elements within the same document. Frequently applied in business intelligence to intuitively identify patterns and infer relationships between concepts that are not inherently linked.
Corpus: The complete collection of linguistic data intended for analysis — a well-structured compilation of documents that should accurately reflect the types of texts an NLP system will encounter in real-world use.
Critical AI: A method of analyzing artificial intelligence through a lens that emphasizes thoughtful evaluation and critique to comprehend and question both current and past frameworks within AI.
Crossover: In genetic algorithms and evolutionary computation, a genetic mechanism employed to merge the hereditary data of two parent solutions to produce novel offspring, resembling the recombination process in sexual reproduction among biological entities.
Custom/Domain Language Model: A model designed explicitly for a particular organization or sector, such as the insurance industry.

D

Darkforest: A computer Go application created by Facebook, utilizing deep learning methodologies with a convolutional neural network. Its improved iteration, Darkforest2, integrates its predecessor's strategies with Monte Carlo tree search (MCTS). Following this upgrade, the system is referred to as Darkfmcts3.
Dartmouth Workshop: The Dartmouth Summer Research Project on Artificial Intelligence — a 1956 summer workshop widely regarded as the foundational milestone for AI as a discipline.
Data Augmentation: Methodologies employed to expand the volume of data in data analysis. It aids in minimizing overfitting when training a machine learning model.
Data Discovery: The process of discovering valuable data insights and delivering them to the appropriate users at the right time.
Data Drift: Takes place when the distribution of input data evolves, also referred to as covariate shift.
Data Extraction: The method of gathering or obtaining various types of raw data from multiple sources, many of which may be disorganized or entirely unstructured.
Data Fusion: The procedure of merging various data sources to generate more coherent, precise, and valuable information than that offered by any single data source.
Data Ingestion: The method of acquiring diverse data from various sources, reorganizing it, and converting it into a standardized format or repository to increase usability.
Data Integration: The method of merging data stored in distinct sources and presenting users with a consolidated perspective of them. This procedure gains importance when two comparable firms must unify their databases, and as the scale of data expands and the necessity to share existing information surges.
Data Labeling: A method of labeling data to help machines identify objects. Additional information is applied to different data formats — such as text, audio, images, and video — to generate metadata used for training AI models.
Datalog: A declarative logic-based programming language that forms a subset of Prolog. Frequently employed as a query language for deductive database systems, and gaining renewed utility in domains such as data fusion, knowledge extraction, networking, software analysis, cybersecurity, and cloud-based computing.
Data Mining: The practice of analyzing extensive data sets to uncover patterns that improve models or address challenges.
Data Scarcity: The absence of sufficient data that could potentially fulfill the system's requirement to increase the precision of predictive analytics.
Data Science: A multidisciplinary field of technology that leverages algorithms and methodologies to collect and examine vast amounts of data, revealing patterns and insights that guide business decisions.
Data Set: An aggregation of information. Typically, a dataset corresponds to the contents of a singular database table or a single statistical data matrix, where each column signifies a specific variable and every row represents a distinct entity. Each individual value is referred to as a datum.
Data Warehouse: A framework employed for reporting and data examination. Data warehouses serve as centralized repositories that consolidate information from multiple distinct sources, retaining both present and past data within a unified location.
Decision Boundary: For ANNs or perceptrons utilizing backpropagation, the nature of the decision boundary is influenced by the quantity of hidden layers. Without hidden layers, the network is restricted to learning only linear problems. With a single hidden layer, it can approximate any continuous function over compact subsets, enabling it to represent an arbitrary decision boundary.
Decision Support System (DSS): An information system designed to aid business or organizational decision-making processes. DSSs assist management at operational and strategic planning levels by facilitating decision-making for issues that evolve quickly and are not easily predefined, such as unstructured or semi-structured decision challenges.
Decision Theory: The examination of the logic behind an agent's decision-making process. Divided into normative decision theory (which provides guidance on making optimal choices) and descriptive decision theory (which investigates how real-world agents actually arrive at decisions).
Decision Tree Learning: Employs a decision tree as a forecasting framework to map observations regarding an entity (depicted in the branches) to determine the entity's target attribute (represented in the leaves). Among the predictive modeling methodologies utilized in statistics, data mining, and machine learning.
Declarative Programming: A coding paradigm that conveys the logic of a computation without specifying its execution sequence.
Deductive Classifier: A category of artificial intelligence reasoning engine that processes as input a collection of statements in a frame-based language concerning a domain like medical studies or molecular biology, including names of categories, subcategories, attributes, and constraints on permissible values.
Deep Blue: A computer system designed by IBM for playing chess. Recognized as the first chess-playing machine to secure both a game and an entire match victory against a reigning world champion under standard time constraints.
Deep Learning: A branch of AI that mimics the human brain's structure by analyzing how it organizes and processes information to make decisions. Rather than depending on an algorithm designed for a single specific task, this subset of machine learning can extract insights from unstructured data without human oversight.
DeepMind Technologies: A British AI firm established in September 2010, now a subsidiary of Alphabet Inc., headquartered in London. Acquired by Google in 2014, the company gained global attention in 2016 when its AlphaGo system triumphed over world champion Go player Lee Sedol. A more advanced system, AlphaZero, surpassed the strongest existing programs in Go, chess, and shogi after only a few days of self-play using reinforcement learning.
Default Logic: A non-monotonic reasoning framework introduced by Raymond Reiter to systematize inference based on default presumptions.
Density-Based Spatial Clustering of Applications With Noise (DBSCAN): A grouping algorithm introduced by Martin Ester, Hans-Peter Kriegel, Jörg Sander, and Xiaowei Xu in 1996.
Description Logic (DL): A collection of structured knowledge representation languages offering greater expressiveness than propositional logic while remaining less expressive than first-order logic. Fundamental reasoning tasks in DLs are typically decidable, and efficient computational procedures have been developed and implemented to address these tasks.
Developmental Robotics (DevRob): A scientific discipline focused on investigating the developmental processes, frameworks, and limitations that support continuous and unrestricted acquisition of new abilities and knowledge in physically embedded machines.
Diagnosis: Focused on creating algorithms and methodologies capable of verifying whether a system's behavior is accurate. If the system operates incorrectly, the algorithm should identify, with maximum precision, the malfunctioning component and the nature of the fault.
Dialogue System: A computational system designed to engage in structured interactions with a human. Conversational systems have utilized text, speech, visuals, touch-based inputs, gestures, and various other modalities for communication on both input and output channels.
Did You Mean (DYM): A natural language processing feature utilized in search applications to detect misspellings in a query or propose alternative queries that may yield relevant results within the search database.
Diffusion Model: In machine learning, also referred to as diffusion probabilistic models or score-based generative models, a category of latent variable models that function as Markov chains trained via variational inference. In computer vision, this entails training a neural network to restore images degraded by Gaussian noise by learning to reverse the diffusion process.
Dijkstra's Algorithm: An algorithm designed to determine the shortest routes between vertices in a weighted graph, which can symbolize transportation networks and other systems.
Dimensionality Reduction: The procedure of minimizing the quantity of random variables being analyzed by deriving a set of key variables. It can be categorized into feature selection and feature extraction.
Disambiguation: Also known as word-sense disambiguation, the method of eliminating ambiguity in terms that have multiple meanings, preventing misinterpretation of the same sequence of text.
Discrete System: Any mechanism possessing a countable set of states, contrasted with continuous (analog) systems. Since discrete systems have a countable number of states, they can be precisely described through mathematical models. A computer functions as a finite-state machine and can be considered a discrete system.
Distributed Artificial Intelligence (DAI): A branch of artificial intelligence research focused on creating decentralized solutions for various problems.
Domain Knowledge: The knowledge and proficiency an organization has accumulated in a specific field.
Double Descent: A statistical and machine learning phenomenon in which both a model with few parameters and one with an exceptionally high number of parameters exhibit low test error, whereas a model with a parameter count roughly equivalent to the number of training data points incurs high error. This challenges traditional assumptions regarding overfitting in classical machine learning.
Dropout: A normalization method for minimizing overfitting in ANNs by inhibiting complex co-adaptations on training datasets.
Dynamic Epistemic Logic (DEL): A formal system addressing the dynamics of knowledge and information modification. Generally, DEL concentrates on scenarios with multiple agents and examines how their awareness evolves in response to occurring events.

E

Eager Learning: A training approach where the system endeavors to formulate a broad, input-agnostic target function during the learning phase, in contrast to lazy learning, where the process of generalization beyond the training dataset is postponed until the system receives a query.
Early Stopping: A constraint enforcement method frequently employed during the training of a machine learning model using an iterative approach like gradient descent.
Ebert Test: An assessment designed to measure whether a computer-generated synthetic voice can deliver a joke proficiently enough to make listeners laugh. Film critic Roger Ebert introduced this concept at the 2011 TED conference as a challenge for software engineers to develop an artificial voice capable of replicating the nuances of human speech. This evaluation is analogous to the Turing test.
Echo State Network (ESN): A recurrent neural network featuring a sparsely interconnected hidden layer (typically with around 1% connectivity). The connections and weights of the hidden neurons are predetermined and assigned at random. The weights of the output neurons can be adjusted to help the network generate or replicate specific temporal patterns.
Edge Model: A model that incorporates data usually located outside centralized cloud data centers and nearer to local devices or users, such as wearables and Internet of Things (IoT) sensors or controllers.
Embedding: A collection of data frameworks within a large language model (LLM) that represents a body of text, where words are encoded as high-dimensional vectors. This approach increases the efficiency of processing meaning, translation, and the creation of new content.
Embodied Agent: A cognitive agent that engages with its surroundings using a tangible body within that environment. Agents depicted visually with a form — such as a humanoid figure — are also referred to as embodied agents, even though their embodiment may exist solely in a virtual form.
Embodied Cognitive Science: A cross-disciplinary area of study focused on elucidating the mechanisms that drive intelligent conduct, encompassing the simulation of cognitive and biological systems as a unified whole, the establishment of universal principles governing intelligent behavior, and the empirical deployment of robotic entities within regulated settings for experimentation.
Emergent Behavior: Also known as emergence, occurs when an AI system exhibits unexpected or unplanned abilities.
Emotion AI or Affective Computing: AI that evaluates a user's emotional condition through computer vision, voice/audio input, sensors, or software algorithms. It can trigger responses by executing tailored actions to align with the user's mood.
Ensemble Learning: The utilization of multiple machine learning models to achieve superior predictive accuracy compared to what any individual underlying algorithm could accomplish independently.
Entity: Any noun, term, or expression within a document that denotes a concept, individual, object, or abstraction. This category also encompasses quantifiable elements.
Environmental, Social, and Governance (ESG): Originally associated with business and government, relating to an organization's social influence and responsibility. Disclosures in this domain are regulated by a combination of mandatory and voluntary compliance frameworks.
Epoch: In machine learning, especially in the development of ANNs, a single complete pass through the entire training dataset during model training. Smaller models are generally trained for multiple epochs until optimal performance is achieved on the validation dataset, whereas the largest models may undergo training for just one epoch.
Error-Driven Learning: A branch of machine learning focused on determining how an agent should execute actions within an environment to reduce a specified error signal. It represents a form of reinforcement learning.
Ethics of Artificial Intelligence: The branch of technology ethics that pertains specifically to artificial intelligence.
ETL or Extract, Transform, Load: Entity extraction is a natural language processing capability that detects and identifies significant entities within a document.
Evolutionary Algorithm (EA): A branch of evolutionary computation — a broad population-based metaheuristic optimization method. An evolutionary algorithm employs principles derived from biological evolution, including reproduction, mutation, recombination, and selection. Potential solutions to the optimization problem function as individuals within a population, while a fitness function evaluates their quality.
Evolutionary Computation: A class of algorithms for worldwide optimization influenced by natural evolution, as well as the domain of AI and soft computing that examines these methods. They constitute a group of population-based, trial-and-error problem-solving techniques with a metaheuristic or probabilistic optimization nature.
Evolving Classification Function (ECF): Adaptive classification functions utilized for categorization and grouping within machine learning and AI, commonly applied in data stream analysis tasks within fluid and evolving environments.
Existential Risk: The conjecture that significant advancements in artificial general intelligence (AGI) might eventually lead to the extinction of humanity or another irreversible worldwide disaster.
Expert System: A computational system that replicates the decision-making capabilities of a human specialist. Expert systems are engineered to tackle intricate problems by logically processing extensive knowledge bases, primarily represented as if-then rules rather than traditional procedural programming.
Explainable AI/Explainability: An AI methodology in which the functioning of its algorithms is transparent and comprehensible to humans. Unlike black-box AI, this approach provides visibility into the decision-making process and the rationale behind its outcomes.
Extraction or Keyphrase Extraction: Several terms that capture the core concepts and fundamental meaning of the text within documents.
Extractive Summarization: Extracts key details from a text and clusters related fragments to create a succinct summary.

F

Fast-And-Frugal Trees: A form of decision tree used for classification. Fast-and-frugal trees serve as decision-making instruments that function as lexicographic classifiers and, when necessary, assign an action or choice to each category or class.
Feature: A distinct quantifiable attribute or trait of a phenomenon. In computer vision and image analysis, a feature represents a fragment of data regarding an image's content, usually indicating whether a specific area of the image possesses particular characteristics.
Feature Extraction: In machine learning, pattern recognition, and image analysis, begins with an original collection of recorded data and generates transformed values (features) designed to be insightful and non-redundant. This process aids in the subsequent phases of learning and generalization and, in certain instances, supports human interpretability.
Feature Learning: In machine learning, also known as representation learning, a collection of methods enabling a system to autonomously identify the representations essential for feature recognition or categorization from unprocessed data. It eliminates the need for manual feature engineering, allowing the machine to simultaneously acquire the features and apply them to execute a particular task.
Feature Selection: Also referred to as variable selection, attribute selection, or variable subset selection, the procedure of identifying a subset of pertinent features (variables, predictors) for incorporation in model development.
Federated Learning: A machine learning approach that supports the training of models on multiple devices using distributed data, thereby aiding in safeguarding the privacy of individual users and their information.
Few-Shot Learning: Unlike conventional models that rely on extensive training datasets, few-shot learning requires only a limited number of training examples to generalize effectively and generate meaningful results.
Fine-Tuned Model: A model tailored to a particular domain or classification of information, such as a subject, sector, or set of challenges.
Fine-Tuning: Enhancing a previously trained model by further refining it with new data tailored to a specific context or task.
First-Order Logic: A set of formal frameworks utilized in mathematics, philosophy, linguistics, and computer science. First-order logic employs quantified variables over non-logical entities and permits the formation of statements containing variables, distinguishing it from propositional logic which lacks quantifiers and relational expressions.
Fluent: A state or attribute that varies with the passing of time. In logical methodologies for reasoning about actions, fluents can be expressed in first-order logic through predicates that include a time-dependent argument.
Forward Chaining: One of the two primary reasoning techniques employed in an inference engine. Forward chaining begins with known information and applies inference rules to derive additional facts until a target outcome is achieved. It is the inverse method of backward chaining and is widely used in expert systems, business applications, and production rule frameworks.
Foundational Model: A core model serving as the foundation for a solution set, generally pre-trained on vast datasets through self-supervised machine learning. Other models or applications are built upon foundational models or adapted into fine-tuned, context-specific variants. Examples include BERT, GPT-n, Llama, and DALL-E.
Frame: A data structure in artificial intelligence designed to segment knowledge into smaller substructures by depicting typical scenarios. Frames serve as the fundamental data representation format in AI frame-based languages.
Frame Language: A methodology for representing knowledge in AI. Frames are organized as ontologies comprising groups and subgroups of conceptual frames. They resemble class hierarchies in object-oriented programming languages, though their core objectives differ — frames emphasize clear and intuitive knowledge representation, whereas objects prioritize encapsulation.
Frame Problem: The challenge of identifying suitable sets of axioms to provide a comprehensive and functional representation of a robotic environment.
Friendly Artificial Intelligence: A theoretical artificial general intelligence designed to benefit humanity. It falls within the domain of AI ethics and is closely linked to machine ethics, focusing on the practical implementation of beneficial AI behavior and ensuring it remains appropriately regulated.
F-Score (F-Measure, F1 Measure): Represents the harmonic mean of a system's precision and recall metrics. Computed using the formula: 2 × [(Precision × Recall) / (Precision + Recall)]. The F2 measure places greater emphasis on recall, while the F0.5 measure gives more weight to precision.
Futures Studies: The exploration of theorizing potential, likely, and desirable futures, along with the perspectives and narratives that shape them.
Fuzzy Control System: A regulatory mechanism utilizing fuzzy logic or a computational framework that interprets analog input magnitudes through logical variables that assume continuous values ranging from 0 to 1, as opposed to traditional binary logic which functions with discrete states of either 1 or 0.
Fuzzy Logic: A basic variant of many-valued logic where the truth values of variables can assume any degree of truthfulness represented by any real number within the inclusive range of 0 (Completely False) to 1 (Completely True). Utilized to manage the notion of partial truth, contrasting with Boolean logic in which variables can only take on integer values of 0 or 1.
Fuzzy Rule: A guideline applied in fuzzy logic systems to deduce an outcome based on input parameters.
Fuzzy Set: In fuzzy set theory, elements have a gradual evaluation of membership within a set, represented using a membership function that takes values within the real unit interval [0,1]. This extends classical sets, where inclusion of elements is evaluated in a binary manner. Applicable in fields where data is incomplete or imprecise, such as bioinformatics.

G

Game Theory: The analysis of mathematical frameworks that model strategic interactions among rational decision-makers.
General Game Playing (GGP): Involves creating AI systems capable of executing and competently engaging in multiple games.
Generalization: The notion that humans, other animals, and artificial neural networks apply prior knowledge to current learning scenarios when the circumstances are perceived as analogous.
Generalization Error: In supervised learning tasks within machine learning and statistical learning theory, also referred to as out-of-sample error or risk, this quantifies how precisely a learning algorithm can forecast results for data it has not encountered before.
Generalized Model: A model that is not tailored to particular applications or specific types of information.
Generative Adversarial Network (GAN): A category of machine learning models where two neural networks compete against one another within a zero-sum game structure.
Generative AI (GenAI): A form of technology that leverages artificial intelligence to produce content such as text, video, code, and images. A generative AI system is trained on vast amounts of data, allowing it to identify patterns for generating new material.
Generative Pretrained Transformer (GPT): A substantial language model built upon the transformer framework that produces text. Initially undergoes pretraining to forecast the next token in sequences. GPT models are capable of producing text resembling human writing and are typically further refined through reinforcement learning guided by human input.
Generative Summarization: Leveraging large language model capabilities to process text-based inputs such as extended conversations, emails, reports, contracts, and policies, extracting key information and condensing it into essential summaries. The process involves utilizing pretrained language models and contextual comprehension to generate brief, precise, and pertinent summaries.
Genetic Algorithm (GA): A metaheuristic based on the concept of natural selection falling under the broader category of evolutionary algorithms. Genetic algorithms are frequently employed to produce optimal solutions for optimization and search issues, utilizing bio-inspired techniques such as mutation, crossover, and selection.
Genetic Operator: An operator utilized in genetic algorithms to steer the algorithm toward a solution for a specific problem. There are three primary types — mutation, crossover, and selection — which must function together for the algorithm to be effective.
Glowworm Swarm Optimization: A swarm intelligence optimization algorithm inspired by the behavior of glowworms, also referred to as fireflies or lightning bugs.
Gradient Boosting: A machine learning method based on boosting in a functional space, where the objective is pseudo-residuals rather than residuals as in conventional boosting.
Graph (Abstract Data Type): An abstract data structure designed to represent the concepts of undirected and directed graphs from mathematics, particularly within the domain of graph theory.
Graph (Discrete Mathematics): In mathematics, particularly in graph theory, a structure consisting of a collection of objects where certain pairs of these objects are connected. The objects are referred to as vertices (nodes or points) and each connected pair of vertices is called an edge (arc or line).
Graph Database (GDB): A database that employs graph structures for semantic queries, using nodes, edges, and attributes to represent and store information. These relationships allow direct linking of data within the store, often allowing retrieval with a single operation. Particularly useful for highly interconnected data.
Graph Theory: The analysis of graphs, which are mathematical constructs used to represent pairwise connections between entities.
Graph Traversal: The procedure of visiting, examining, and/or modifying each node in a graph. These traversals are categorized by the sequence in which the nodes are visited. Tree traversal is a specific instance of graph traversal.
Grounding: The capability of generative systems to trace the factual content within a generated output or completion. It connects generative applications to accessible factual sources, such as documents or knowledge repositories, either by providing a direct citation or by searching for additional references.
Guardrails: Constraints and guidelines imposed on AI systems to ensure they process data responsibly and do not produce unethical content.

H

Hallucination: An inaccurate response from an AI system or misleading information in an output that is presented as if it were true.
Hallucitations: Fictitious data consisting of invented, incorrect, or misleading references or sources mistakenly presented as factual within generated content.
Heuristic: A method developed to solve a problem more rapidly when traditional approaches are too sluggish, or to find an approximate solution when traditional methods cannot identify an exact one. Accomplished by sacrificing optimality, completeness, accuracy, or precision in favor of speed. A heuristic function ranks options in search algorithms at each decision point based on the available data to determine which path to pursue.
Hidden Layer: A layer of neurons in an artificial neural network that is neither an input layer nor an output layer.
Human-Centered Perspective: A human-focused viewpoint that envisions AI systems collaborating with individuals and enhancing human abilities. Humans should always maintain a primary role in education, and AI should not serve as a substitute for teachers.
Hybrid AI: Any artificial intelligence system that integrates multiple AI approaches. In natural language processing, this typically involves utilizing both symbolic reasoning and machine learning techniques within a single workflow.
Hyper-Heuristic: A heuristic search approach that aims to automate the process of choosing, merging, creating, or modifying multiple simpler heuristics to effectively resolve computational search issues, often through the integration of machine learning methods.
Hyperparameter: A variable or setting that influences how an AI model learns. It is typically configured manually outside the model.
Hyperparameter Optimization: The procedure of selecting an ideal set of hyperparameters for a learning algorithm.
Hyperplane: A decision boundary in machine learning classifiers that divides the input space into multiple regions, with each region representing a distinct class label.

I

IEEE Computational Intelligence Society: A professional organization within the Institute of Electrical and Electronics Engineers (IEEE) concentrating on the principles, design, implementation, and advancement of biologically and linguistically inspired computational frameworks, emphasizing neural networks, connectionist systems, genetic algorithms, evolutionary programming, fuzzy systems, and hybrid intelligent systems.
Image Recognition: The process of detecting and identifying objects, individuals, locations, or text within an image or video.
Incremental Learning: A machine learning approach in which input data is progressively utilized to expand the current model's understanding. It signifies a flexible method of supervised and unsupervised learning that can be employed when training data is provided incrementally or when its volume exceeds system memory capacity.
Inference Engine: A module within an expert system that utilizes logical principles to analyze the knowledge base and infer new or supplementary information.
Information Integration (II): The integration of data from several sources with varying conceptual, contextual, and typographical formats. Employed in data mining and the consolidation of information from unstructured or semi-structured sources.
Information Processing Language (IPL): A programming language designed with features to assist in creating programs that carry out basic problem-solving tasks, such as lists, dynamic memory allocation, data types, recursion, functions as parameters, generators, and cooperative multitasking. IPL introduced the concept of list processing.
Insight Engine: Also known as cognitive search or enterprise knowledge discovery, employs relevance-based techniques to interpret, uncover, structure, and examine data. It merges search functionality with AI to supply information for users and datasets for machines, with the primary objective of delivering timely data that yields actionable insights.
Intelligence Amplification (IA): The efficient application of information technology in enhancing human intellect.
Intelligence Explosion: A potential consequence of humanity creating artificial general intelligence. AGI would possess the ability to recursively improve itself, resulting in the swift rise of artificial superintelligence at the moment of the technological singularity.
Intelligent Agent (IA): A self-sufficient entity that performs actions focusing its efforts on reaching objectives. It functions within an environment by perceiving through sensors and responding with actuators. Intelligent agents can also acquire knowledge or use existing information to fulfill their objectives, ranging from very basic to highly intricate.
Intelligent Control: A category of control methods that employ different artificial intelligence computing strategies such as neural networks, Bayesian probability, fuzzy logic, machine learning, reinforcement learning, evolutionary computation, and genetic algorithms.
Intelligent Document Processing (IDP)/Intelligent Document Extraction and Processing (IDEP): The capability to automatically interpret and transform unstructured and semi-structured data, recognize relevant information, extract it, and utilize it through automated workflows. Frequently a foundational technology for Robotic Process Automation (RPA) operations.
Intelligent Personal Assistant: A software agent capable of carrying out tasks or providing services for a user based on spoken instructions. Some virtual assistants have the ability to understand spoken language and reply with synthesized voices, allowing users to ask questions, control smart home devices, and manage tasks like email, to-do lists, and calendars with verbal instructions.
Intelligent Tutoring Systems (ITS): A computational platform or digital educational environment that provides immediate and personalized feedback to learners. An ITS can utilize rule-based AI or employ machine learning behind the scenes, facilitating adaptive learning.
Interpretable Machine Learning (IML): Also known as interpretable AI, refers to the development of models that are naturally understandable, offering built-in explanations for their decisions. This method is favored over explainable machine learning because it provides understanding from the outset rather than attempting to clarify opaque models afterward.
Interpretation: An attribution of meaning to the symbols of a formal language. The broader study of the interpretations of formal languages is referred to as formal semantics.
Intrinsic Motivation: An intelligent agent is internally driven to act as if the informational value of the outcome of the action serves as the motivating element. A common form of intrinsic motivation is the pursuit of novel or unexpected scenarios — unlike extrinsic motivations such as seeking food. Artificial agents guided by intrinsic motivation exhibit behaviors similar to exploration and curiosity.
Issue Tree: A visual representation of a question that breaks it down into its various elements vertically, with increasing detail as it moves to the right. Helpful in problem-solving to pinpoint the underlying causes of an issue and to discover possible solutions.

J

Junction Tree Algorithm: A technique employed in machine learning to derive marginalization in general graphs. Essentially, it involves executing belief propagation on a transformed graph known as a junction tree. The graph is termed a tree due to its branching structure, where the nodes of variables represent the branches.

K

Kernel Method: In machine learning, a category of algorithms used for pattern recognition, with the most well-known being the support vector machine (SVM). The primary goal of pattern recognition is to identify and examine general types of relationships within datasets, including cluster analysis, rankings, principal components, correlations, and classifications.
KL-ONE: A renowned knowledge representation framework in the tradition of semantic networks and frames (a frame-based language). The system seeks to address semantic ambiguity in semantic network models and to clearly represent conceptual knowledge as an organized inheritance structure.
K-Means Clustering: A technique of vector quantization, initially derived from signal processing, that seeks to divide n data points into k groups, with each data point assigned to one of the groups. The closest average (cluster center or centroid) acts as the representative of the cluster.
K-Nearest Neighbors: A non-parametric supervised learning technique initially created by Evelyn Fix and Joseph Hodges in 1951 and later improved by Thomas Cover. Utilized for both classification and regression tasks.
Knowledge Acquisition: The procedure used to establish the rules and ontologies needed for a knowledge-based system. Originally employed in relation to expert systems to describe the initial steps of consulting domain specialists and recording their expertise through rules, objects, and frame-based ontologies.
Knowledge-Based AI: Knowledge-based systems (KBS) are a type of AI built to encapsulate the expertise of human specialists, aiding in decision-making and resolving problems.
Knowledge-Based System: A software application that uses reasoning and a knowledge repository to address intricate issues. The shared principle between all knowledge-based systems is the effort to represent knowledge clearly and a reasoning mechanism that supports the generation of new knowledge.
Knowledge Distillation: The method of transferring knowledge from a larger machine learning model to a more compact one.
Knowledge Engineering (KE): A technique for helping computers emulate human-like understanding. Knowledge engineers embed reasoning into knowledge-based systems by gathering, structuring, and incorporating general or specialized expertise into a framework.
Knowledge Extraction: The process of generating knowledge from structured relational databases, XML, and unstructured sources such as text, documents, and images. The resulting knowledge must be in a machine-readable and interpretable format to support reasoning.
Knowledge Graph: A network of interconnected concepts whose significance lies in its capacity to accurately depict a segment of reality. Each concept is linked to at least one other, with the nature of these connections categorized into different types. The meaning of each concept is defined by its relationships, making every node representative of its concept solely based on its placement within the graph.
Knowledge Interchange Format (KIF): A programming language created to allow systems to exchange and reuse information from knowledge-based systems. Designed for the exchange of knowledge between systems that may use different languages, formalisms, platforms, and so on.
Knowledge Model: A method for developing a machine-readable representation of knowledge or guidelines related to a language, domain, or set of processes. Structured in a data format that allows the information to be stored in a database and understood by software.
Knowledge Representation and Reasoning (KR² or KR&R): The branch of artificial intelligence focused on representing information about the world in a format that a computer system can use to tackle intricate tasks, such as diagnosing medical conditions or engaging in conversations in a natural language. Examples of knowledge representation frameworks include semantic networks, system architectures, frames, rules, and ontologies.

L

LangOps or Language Operations: The procedures and methodologies that facilitate the development, formulation, evaluation, implementation, and continuous refinement of linguistic models and natural language solutions.
Language Data: Linguistic data consists of words and constitutes a type of unstructured information. This qualitative data, also called textual data, pertains to the written and spoken expressions in a language.
Language Model: A stochastic model that processes natural language.
Large Language Model (LLM): An AI system trained on vast quantities of text, allowing it to comprehend language and produce human-like text.
Lazy Learning: In machine learning, a technique where the generalization of the training data is theoretically postponed until a query is presented to the system, unlike eager learning where the system attempts to generalize the training data prior to receiving any queries.
Lemma: The root form of a word that serves as the representation for all its conjugated or declined variations.
Lexicon: Awareness of all potential interpretations of words within their appropriate context, essential for accurately analyzing textual content with a high degree of precision.
Limited Memory: A category of AI systems that gathers information from real-time events and retains it in a database to improve predictions.
Linked Data: Interconnected data that indicates whether a recognizable repository of information is associated with another. Commonly utilized as a standardized reference — for example, a knowledge graph where each concept or node is connected to its corresponding entry on Wikipedia.
Lisp Programming Language (LISP): A group of programming languages with an extensive history and a unique, completely parenthesized prefix syntax.
Logic Programming: A kind of programming paradigm primarily grounded in formal logic. Any program created in a logic programming language consists of a collection of statements in logical form, representing facts and rules about a specific problem domain. Prominent families include Prolog, answer set programming (ASP), and Datalog.
Long Short-Term Memory (LSTM): A type of artificial recurrent neural network architecture employed in deep learning. In contrast to typical feedforward neural networks, LSTM incorporates feedback connections to function as a universal computer, capable of processing not just individual data points but also complete sequences of data such as speech or video.

M

Machine Learning (ML): A branch of AI that integrates elements of computer science, mathematics, and programming. It emphasizes creating algorithms and models that help machines learn from data and forecast patterns and behaviors without human intervention.
Machine Listening: A broad area of research focused on algorithms and systems for machine-based audio comprehension.
Machine Perception: The ability of a computer system to analyze data in a way that mirrors how humans use their senses to interact with and understand the environment around them.
Machine Vision (MV): The ability of a computer system to analyze data in a manner similar to how humans use their senses to understand and interact with the world. Employed to provide imaging-based automatic inspection and analysis in applications such as automated examination, process management, and robotic navigation, primarily in industrial settings.
Markov Chain: A probabilistic model that represents a series of potential events, where the likelihood of each event is determined solely by the condition reached in the immediately preceding event.
Markov Decision Process (MDP): A discrete-time probabilistic control process offering a mathematical structure for representing decision-making in scenarios where results are partially random and partially influenced by the choices of a decision-maker. Valuable for analyzing optimization problems addressed through dynamic programming and reinforcement learning.
Mathematical Optimization: A field providing a mathematical framework for modeling decision-making in situations where outcomes are partly random and partly shaped by the decisions of a chooser. Useful for examining optimization challenges tackled through dynamic programming and reinforcement learning.
Mechanism Design: A domain within economics and game theory that adopts an engineering perspective to create economic mechanisms or incentives aimed at achieving specific goals in strategic environments. Also referred to as reverse game theory, with wide-ranging uses in economics, politics, markets, auctions, voting methods, and networked systems.
Mechatronics: A cross-disciplinary field of engineering that concentrates on the design and development of both electrical and mechanical systems. It incorporates elements of robotics, electronics, computing, telecommunications, systems, control, and product engineering.
Metabolic Network Reconstruction and Simulation: Provides a comprehensive understanding of the molecular processes within a specific organism, linking the genome to molecular physiology.
Metacontext and Metaprompt: Core guidelines on structuring the training process to shape the model's expected behavior.
Metadata: Information that characterizes or offers details regarding other data.
Metaheuristic: An advanced method or strategy intended to discover, create, or choose a heuristic. The partial search algorithm of the heuristic aids in solving problems with insufficient or imperfect data or constrained computational resources. Metaheuristics explore a set of solutions that is too extensive to be fully sampled.
Model: A machine learning model is the resulting construct generated once an ML algorithm has processed the provided sample data during its training stage. This model is subsequently utilized by the algorithm in deployment to interpret text in NLP scenarios and deliver insights and/or forecasts.
Model Checking: Also known as property verification, involves thoroughly and automatically assessing whether a specific model of a system adheres to a specified criterion. A method for autonomously validating the correctness attributes of finite-state systems.
Model Drift: The deterioration of a model's predictive accuracy due to shifts in real-world conditions. This decline occurs for various reasons, such as alterations in the digital landscape and subsequent changes in the relationships between variables.
Model Parameter: Adjustable factors within the model that are established through training data. They represent the configured or optimized internal variables whose values can be inferred from data, essential for the model when generating predictions.
Modus Ponens: In propositional logic, an inference rule that can be expressed as: 'P implies Q, and P is affirmed to be true, so Q must also be true.'
Modus Tollens: In propositional logic, a valid form of argument and an inference rule applying the general principle that if a statement is true, then its contrapositive must also be true. It asserts that reasoning from 'P implies Q' to 'the negation of Q, implying the negation of P' is valid.
Monte Carlo Tree Search: A heuristic search technique used for certain types of decision-making processes.
Morphological Analysis: Decomposing a problem that has numerous established solutions into its fundamental components or simplest forms to achieve a deeper understanding. Applied in broad problem-solving, language studies, and biological sciences.
Multi-Agent System (MAS): A computerized framework consisting of several interacting intelligent agents. Multi-agent systems can tackle challenges that are too complex or unachievable for a single agent or a unified system. Intelligence may encompass systematic, functional, procedural methods, algorithmic exploration, or reinforcement learning.
Multilayer Perceptron (MLP): In deep learning, a contemporary feedforward neural network made up of fully connected neurons with nonlinear activation functions arranged in layers, recognized for its ability to differentiate data that cannot be separated linearly.
Multimodal Models and Modalities: Linguistic models trained to comprehend and process various data modalities, including text, visuals, sound, and other formats, leading to better performance within a broader spectrum of tasks.
Multi-Swarm Optimization: A variation of particle swarm optimization (PSO) that utilizes multiple sub-swarms rather than a single standard swarm. Each sub-swarm concentrates on a particular area, while a specific diversification technique determines when and where to deploy the sub-swarms. Particularly suited for optimization in multimodal issues where several local optima are present.
Multitask Prompt Tuning (MPT): A method that structures a prompt as a modifiable variable to enable repeated prompts where only the variable is adjusted.
Mutation: A genetic operator employed to preserve genetic diversity between generations of a population in a genetic algorithm. Mutation modifies one or more gene values in a chromosome from its original state, allowing the genetic algorithm to potentially find a better solution. Mutation takes place during evolution based on a user-defined mutation probability, which should be kept low.
MYCIN: An early backward chaining expert system that utilized artificial intelligence to identify bacteria responsible for serious infections, such as bacteremia and meningitis, and to suggest antibiotics with dosage adjusted according to the patient's body weight. The name originated from antibiotics, as many of them end with the suffix -mycin. MYCIN was also employed to diagnose blood clotting disorders.

N

Naive Bayes Classifier: In machine learning, a group of straightforward probabilistic classifiers that rely on applying Bayes' theorem with strong assumptions of independence among the features.
Naive Semantics: A method utilized in computer science for encoding fundamental knowledge about a particular domain. Applied toward understanding the meaning of natural language statements in AI systems. In a broader context, the term describes the application of a restricted set of commonly recognized knowledge about a specific area of the world.
Name Binding: In programming languages, the linking of entities (either data or code) with identifiers. An identifier linked to an object is said to reference that object. Name-object bindings are implemented as a service and notation for the programmer. Binding is closely related to scoping.
Named-Entity Recognition (NER): A subset of information extraction focused on identifying and categorizing named entity references in unstructured text into predetermined categories, such as personal names, organizations, places, medical codes, temporal expressions, quantities, financial values, percentages, and more.
Named Graph: A fundamental idea of Semantic Web architecture, where a collection of Resource Description Framework (RDF) statements is identified using a URI, enabling the description of that collection with additional information such as context, provenance, or other types of metadata.
Natural Language Generation (NLG): Systems that autonomously transform organized data stored in a database, application, or real-time stream into a text-based narrative. This ability expands user accessibility by allowing information to be read or heard, thereby improving understanding.
Natural Language Processing (NLP): A branch of AI that allows computers to interpret and comprehend both spoken and written human language. NLP powers functionalities such as speech and text recognition on various devices.
Natural Language Programming: A method of programming supported by ontology, expressed through natural-language sentences such as English.
Natural Language Query (NLQ): A natural language input consisting solely of words and expressions as they appear in verbal communication, excluding any non-linguistic symbols or characters.
Natural Language Technology (NLT): A specialized branch of linguistics, computer science, and AI that focuses on natural language processing, natural language understanding, and natural language generation.
Natural Language Understanding (NLU): A branch of natural language processing that concentrates on the genuine machine interpretation of examined and structured unstructured linguistic data. The process is facilitated through semantics.
Network Motif: Recurring and statistically meaningful subgraphs or patterns identified across all types of networks — biological, social, technological, and others — which can be depicted as graphs.
Neural Machine Translation (NMT): A method of machine translation that employs a vast artificial neural network to estimate the probability of a word sequence, usually representing complete sentences within a unified model.
Neural Network: A deep learning method modeled after the structure of the human brain. These networks rely on vast data sets to carry out computations and generate outputs, supporting capabilities such as speech and image recognition.
Neural Turing Machine (NTM): A recurrent neural network architecture that merges the flexible pattern recognition abilities of neural networks with the computational strength of programmable systems. An NTM features a neural network controller linked to external memory resources which it engages with via attention mechanisms, allowing it to deduce basic algorithms like copying, sorting, and associative recall solely from examples.
Neurocybernetics: A direct communication channel between an augmented or wired brain and an external device, supporting two-way information exchange. Frequently focused on studying, mapping, supporting, enhancing, or restoring human cognitive or sensory-motor abilities.
Neuro-Fuzzy: Mergers of artificial neural networks and fuzzy reasoning.
Neuromorphic Engineering: A concept referring to the application of very-large-scale integration (VLSI) systems incorporating electronic analog circuits to replicate neuro-biological structures found in the nervous system. The term has been applied to describe analog, digital, hybrid analog/digital VLSI, and software systems that model neural systems for tasks such as perception, motor regulation, or multisensory coordination.
Node: A fundamental component of a data structure, such as a linked list or tree structure. Nodes store data and may also connect to other nodes. Connections between nodes form the structure of data organizations like trees, graphs, and linked lists.

O

Ontology: In AI and computer science, a formal representation of knowledge as a set of concepts within a domain, and the relationships between those concepts. Ontologies are used to reason about the properties of a domain, and they define a common vocabulary for researchers who need to share information in a specific field.
Open-World Assumption: A reasoning paradigm in which a lack of information about a statement does not imply the statement is false. Contrasts with the closed-world assumption used in most traditional databases, where anything not known to be true is assumed to be false.
Optimization: The process of adjusting a system or algorithm to achieve the best possible performance according to a defined objective function. In machine learning, optimization typically involves minimizing a loss function to improve model accuracy.
Overfitting: A modeling error that occurs when a machine learning model learns the training data too well, including its noise and random fluctuations, resulting in poor performance on new, unseen data. Overfitting is a central challenge in machine learning and is addressed through techniques such as regularization, dropout, and cross-validation.

P

Parameter: In machine learning, an internal variable of a model whose value is learned from training data. Parameters define the model's behavior and are adjusted during the training process to minimize prediction error.
Parsing: The process of analyzing a string of symbols in natural language or computer languages, conforming to the rules of a formal grammar. In NLP, parsing involves determining the grammatical structure of a sentence.
Perceptron: One of the earliest and simplest types of artificial neural networks, consisting of a single layer of output nodes. A perceptron classifies inputs by computing a weighted sum and applying a threshold function, and it forms the basis for more complex multilayer neural network architectures.
Planning: In artificial intelligence, the process of generating a sequence of actions to achieve a specified goal from an initial state. AI planning systems reason about actions, their preconditions, and their effects to produce coherent action sequences.
Precision: In information retrieval and classification, the fraction of retrieved or predicted positive instances that are actually correct. Precision is calculated as True Positives / (True Positives + False Positives) and measures the exactness of a model's positive predictions.
Prompt: An input provided to a generative AI model, typically in the form of text, that instructs or guides the model's output. Effective prompt design — known as prompt engineering — is a key skill for obtaining useful and accurate responses from large language models.
Prompt Engineering: The practice of designing and refining input prompts to optimize the performance of large language models on specific tasks. Prompt engineering involves crafting instructions, examples, and context to elicit desired responses from AI systems.

Q

Q-Learning: A model-free reinforcement learning algorithm that seeks to learn the value of an action in a particular state. Q-learning allows an agent to learn optimal policies in environments with discrete action spaces without requiring a model of the environment.
Query: In natural language processing and information retrieval, a request for information submitted to a search engine, database, or AI system. Queries can be expressed in natural language, formal query languages, or as structured inputs to retrieval systems.

R

Random Forest: An ensemble machine learning algorithm that constructs a large number of decision trees during training and outputs the class that is the mode of the classes (for classification) or mean prediction (for regression) of the individual trees. Random forests correct for the tendency of decision trees to overfit their training data.
Recall: In information retrieval and classification, the fraction of actual positive instances that are correctly identified by the model. Recall is calculated as True Positives / (True Positives + False Negatives) and measures the completeness of a model's positive predictions.
Recurrent Neural Network (RNN): A class of artificial neural networks where connections between nodes can create cycles, allowing output from some nodes to affect subsequent inputs. RNNs are designed to process sequential data and are used in tasks such as speech recognition, language modeling, and time series prediction.
Regularization: A set of techniques used in machine learning to prevent overfitting by adding a penalty term to the loss function that discourages overly complex models. Common regularization methods include L1 (Lasso), L2 (Ridge), and dropout.
Reinforcement Learning (RL): A branch of machine learning where an agent learns to make decisions by interacting with an environment, receiving rewards for desirable actions and penalties for undesirable ones. The goal is to learn a policy that maximizes cumulative reward over time.
Representation Learning: A set of techniques in machine learning that allow a system to automatically discover useful features or representations from raw data, reducing the need for manual feature engineering. Deep learning architectures are prominent examples of representation learning systems.
Robotics: An interdisciplinary field combining computer science and engineering that focuses on the design, construction, operation, and use of robots. AI plays a central role in modern robotics, enabling robots to perceive their environment, make decisions, and execute actions autonomously.

S

Self-Supervised Learning: A machine learning approach where the model generates its own supervisory signals from the input data rather than relying on externally provided labels. Large language models such as GPT are trained using self-supervised learning by predicting the next word in a sequence.
Semantic Analysis: The process of understanding the meaning of language beyond its syntactic structure, involving the interpretation of words, phrases, and sentences in context. In NLP, semantic analysis enables machines to comprehend intent, sentiment, and relationships between concepts.
Semantic Network: A knowledge representation framework consisting of a network of nodes representing concepts and edges representing the relationships between them. Semantic networks are used in AI to model associations and enable reasoning about conceptual knowledge.
Sentiment Analysis: A natural language processing technique used to identify and categorize opinions expressed in a piece of text, determining whether the writer's attitude toward a particular topic or product is positive, negative, or neutral. Also known as opinion mining.
Speech Recognition: The ability of a machine to identify and translate spoken language into text. Modern speech recognition systems use deep learning models trained on large datasets of audio and transcriptions to achieve high accuracy across diverse accents and environments.
Supervised Learning: A type of machine learning where the model is trained on labeled data — input-output pairs — and learns to map inputs to outputs. The model is then used to predict outputs for new, unseen inputs. Common supervised learning tasks include classification and regression.
Support Vector Machine (SVM): A supervised machine learning algorithm used for classification and regression tasks. An SVM finds the optimal hyperplane that maximally separates classes in a high-dimensional feature space, using kernel functions to handle non-linearly separable data.
Swarm Intelligence: The collective behavior of decentralized, self-organized systems, typically inspired by natural phenomena such as ant colonies, bird flocking, or bee swarms. Swarm intelligence algorithms, such as Particle Swarm Optimization and Ant Colony Optimization, are used to solve complex optimization problems.
Symbolic AI: An approach to artificial intelligence that represents knowledge and reasoning using symbols, rules, and logic — as opposed to statistical or connectionist approaches. Also known as classical AI or GOFAI (Good Old-Fashioned AI), symbolic AI underpins expert systems, knowledge-based systems, and formal logic engines.

T

Text Classification: The process of assigning predefined categories to text documents based on their content, using machine learning or rule-based approaches. Text classification is used in spam detection, sentiment analysis, topic labeling, and content moderation.
Text Mining: The process of extracting structured information and meaningful patterns from unstructured text data using techniques from NLP, machine learning, and statistical analysis. Applications include information extraction, document clustering, and trend analysis.
Tokenization: The process of breaking a text string into smaller units called tokens, which can be words, subwords, or characters. Tokenization is a fundamental preprocessing step in natural language processing and large language model inference.
Transfer Learning: A machine learning technique where a model trained on one task is reused as the starting point for a model on a different but related task. Transfer learning enables the leveraging of pre-trained knowledge to achieve strong performance with less data and training time.
Transformer: A deep learning model architecture based on self-attention mechanisms, introduced in the paper 'Attention Is All You Need' (2017). Transformers are the foundation of state-of-the-art language models including BERT and GPT, and have achieved remarkable performance across NLP, computer vision, and multimodal tasks.
Turing Test: A test of a machine's ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human. Proposed by Alan Turing in his 1950 paper 'Computing Machinery and Intelligence,' the Turing Test remains a landmark concept in AI philosophy and the assessment of machine intelligence.

U

Underfitting: A modeling problem that occurs when a machine learning model is too simple to capture the underlying patterns in the training data, resulting in poor performance on both training and test datasets. Underfitting is the opposite of overfitting and is addressed by increasing model complexity or training duration.
Unsupervised Learning: A type of machine learning where the model is trained on unlabeled data and must discover patterns, structures, or representations on its own without explicit guidance. Common unsupervised learning tasks include clustering, dimensionality reduction, and anomaly detection.

V

Validation: In machine learning, the process of evaluating a model's performance on a held-out dataset (the validation set) during training to tune hyperparameters and detect overfitting. Validation is distinct from testing, which provides a final, unbiased assessment of model performance.
Variational Autoencoder (VAE): A type of generative model and autoencoder that learns a probabilistic latent representation of input data. VAEs can generate new data samples by sampling from the learned latent space and are widely used in image generation, data augmentation, and anomaly detection.
Vector: In machine learning and NLP, a numerical array that represents a data point, word, or document in a multi-dimensional space. Vectors capture semantic relationships, enabling mathematical operations such as similarity computation and clustering.

W

Weight: In artificial neural networks, a numerical parameter associated with a connection between nodes that determines the strength and direction of influence one node has on another. Weights are learned during training through optimization algorithms such as gradient descent and backpropagation.
Word Embedding: A type of word representation in NLP that maps words or phrases to vectors of real numbers in a continuous vector space, capturing semantic and syntactic relationships between words. Prominent word embedding methods include Word2Vec, GloVe, and FastText.
Word Sense Disambiguation (WSD): The process of identifying which meaning of a word is used in context when the word has multiple possible meanings. WSD is a fundamental challenge in NLP that requires understanding linguistic context and semantic relationships.

X

XAI (Explainable Artificial Intelligence): A set of processes and methods that allow human users to comprehend and trust the results and outputs created by machine learning algorithms. XAI is used to describe an AI model, its expected impact, and potential biases, improving transparency and accountability in AI-driven decisions.

Z

Zero-Shot Learning: A machine learning approach where a model is evaluated on tasks or classes that it has never seen during training. Zero-shot learning relies on the model's ability to generalize from training knowledge, often using semantic descriptions or embeddings to bridge the gap between known and unknown classes.
Zero-Shot Prompting: A prompting technique for large language models where the model is given a task description without any examples. The model leverages its pre-trained knowledge to complete the task, demonstrating the breadth of generalization achieved through large-scale pretraining.