227 · A–Z
the removal or disabling of a component of an AI model in order to assess its importance.
an AI model combining neural networks and fuzzy logic to model unclear, nonlinear, and complex problems.
a nonlinear mathematical function applied to the output value of each neuron, enabling the model to learn nonlinear patterns in data.
a visual representation of the activity of the intermediate layers of an artificial neural network, showing which input regions triggered particular filters (e.g., edges, textures, or shapes).
a finite sequence of clearly formulated instructions to be carried out in order to solve a particular task.
a property of an algorithm indicating how much time and resources it uses as the data size changes.
a machine learning method in which several models are trained independently using randomly selected subsets of data, and their results are averaged (for regression) or voted on (for classification).
the predictions of several machine learning models combined into a single shared decision in order to achieve greater accuracy.
a rectangle defining the location of an object in an image, used for the outputs of object detection models and for labeling training data.
the difference between an exact value and its approximation, used in calculations or modeling.
an algorithm used for training multilayer neural networks that, by applying the chain rule, efficiently computes the gradient of the loss function with respect to all network weights, propagating the error signal from the output layer to…
a rectangle of predefined size and aspect ratio, used as a reference point for predicting the locations and sizes of potential objects in object detection tasks.
a computer vision task in which all objects in an image are identified and labeled with pixel-level precision, including objects of the same class.
an artificial neural network trained to simplify (encode) and reconstruct (decode) the simplified information.
the process of processing spoken language and converting it into text format, using AI algorithms and analyzing acoustic properties and language modeling, in order to identify and transcribe spoken words.
the automation of the machine learning model development process, encompassing feature engineering, model selection, and hyperparameter optimization, enabling the creation of AI systems without expert knowledge.
a problem-solving strategy in which a system or agent tries different actions, evaluates their results, and adjusts its choices in pursuit of a particular goal.
a hypothetical artificial intelligence that imitates human cognitive abilities and is capable of performing any intellectual task as well as or better than a human.
a machine learning method that allows linear algorithms to be applied to nonlinear data, using a kernel function to implicitly map the data into a higher-dimensional space.
a computer system composed of several interacting AI agents that operate in a shared environment to pursue individual or common system goals, coordinating their activities and solving complex tasks.
a classification task in which more than one independent category may be assigned to each object simultaneously.
a classification task in which each object must be assigned to only one of three or more possible mutually exclusive categories.
a feedforward neural network with one or more hidden layers, using nonlinear activation functions to model nonlinear patterns in data.
a machine learning strategy in which a single model is trained simultaneously to perform several related tasks, applying the learned information to all of them.
the class with the largest number of examples in the dataset.
1. a data-decoding device that returns data to the state it was in before encoding; 2. a part of an artificial neural network that reconstructs compressed data into the state it was in before encoding.
an AI mechanism that allows different levels of importance to be assigned to elements of the input data during prediction or generation.
the additional training of a pretrained model with domain-specific data, in order to adapt it to specialized tasks.
data characterized by large volume, velocity of creation, and variety, exceeding the capabilities of traditional databases.
a generative AI model trained on a large quantity of text, designed to understand, interpret, and generate human language.
a generative AI model trained on a large quantity of images, designed to understand, interpret, and generate visual data.
the use of AI to make efficient, real-time information-based routing decisions, e.g., in logistics or data networks.
a machine learning system that imitates the mechanisms of the biological immune system, e.g., recognition and adaptation, used for anomaly detection, optimization, or ensuring data security.
a computer system that gathers information about its environment through sensors and performs actions to achieve goals, using methods such as planning, decision-making, or learning.
data generated or used throughout the lifecycle of an AI system's development, training, deployment, and maintenance.
a property of an AI system to withstand adverse events (e.g., technical failures, malicious attacks), adapt to them, and restore operation afterwards.
a property of an AI system to maintain performance and accuracy under unexpected or adverse conditions, e.g., when processing inaccurate or unstructured data.
a property of an AI system to generate consistent and predictable results under similar input conditions, ensuring the system's reliability.
a model's tendency to consistently make the same errors due to excessive model simplification or a mismatch between the training data and the real situation, resulting in a systematic deviation of predictions from true values.
a process for verifying whether an AI system solves the intended problem and meets user needs.
the variability of a model's predictions arising from sensitivity to the training data, when the model overfits to a specific training set and generalizes poorly to new data.
a process for verifying whether an AI system is implemented according to plan and operates as designed.
the ability of computer systems to perform tasks that normally require human intelligence, such as learning, problem-solving, or decision-making.
an information processing structure that processes information by imitating the principles of biological neural networks, composed of interconnected artificial neurons organized into layers.
a reinforcement learning principle in which future rewards are progressively reduced, giving greater weight to near-term results.
a component of generative adversarial networks: a neural network that classifies data as real or generator-produced, thereby encouraging the generator to create more realistic data.
a type of AI model capable of distinguishing and assigning data to different categories according to their features.
a specific part of an image selected for more detailed analysis or processing.
the addition of supplementary information to data by assigning labels or metadata required for training AI models.
the combining of data from different sources into a unified, consistent dataset.
an automated system used for the sequential collection, processing, and transfer of data from a source to a machine learning model.
1. the unauthorized or unintentional disclosure or transfer of confidential data to unauthorized persons or systems; 2. a situation in which training data unintentionally includes information from test or validation sets, artificially…
a smaller portion of a dataset, processed in a single training iteration.
the expansion of a training dataset by creating modified versions of existing data (e.g., by rotating, cropping, or changing the colors of images), in order to improve model training.
a body of data used for training, testing, or evaluating a model.
the integration of data from several different sources into a single, more reliable and comprehensive source of information.
the improvement of data quality by removing noise and duplicate and unnecessary records.
a process in which predefined target values, classes, or categories are assigned to data objects, forming the basis for the training and evaluation of supervised machine learning models.
a recurrent neural network architecture that processes a sequence of elements in both directions (forward and backward), enabling the model to take into account the context before and after each element of the sequence.
a machine learning task in which data is classified into one of two categories.
a computer system that uses a structured knowledge base and a set of "if-then" rules to make decisions in a specific domain.
an activation function that returns positive values unchanged (f(x) = x) and transforms negative values via an exponential function (α(eˣ -- 1)), thereby reducing the dying neurons problem and stabilizing learning.
a reinforcement learning cycle from an initial state to a terminal state, ending when the agent reaches the goal or fails.
a single pass through the entire training dataset during machine learning.
a classification model accuracy metric calculated as the harmonic mean of precision and recall, especially useful in the case of imbalanced data.
a metric for evaluating the quality of generated images, measuring the statistical distance between the feature distributions of generated and real images.
an activation function that probabilistically attenuates input values according to a Gaussian distribution, allowing smoother regulation of the flow of information through the network than the rectified linear unit function.
a model's ability to accurately process new data not analyzed during training.
a deep learning model consisting of two competing neural networks that learn from each other: a generator that creates synthetic data, and a discriminator that classifies it as real or artificial.
a large language model based on the transformer architecture that generates text by sequentially predicting each next token according to the preceding context, using the attention mechanism.
an AI system that generates new content (text, images, audio, program code) based on patterns learned from training data.
a part of generative adversarial networks used to generate new synthetic data.
artificial but realistic-looking content created using AI technologies, such as photos, videos, or audio recordings in which people appear or sound as though they are doing or saying things they did not actually do or say.
a machine learning method using multilayer neural networks that enables automatic extraction of features from raw data and the solving of complex tasks.
an artificial neural network with two or more hidden layers between the input and output layers, processing data hierarchically and identifying features of complex data or nonlinear relationships between data.
a mathematical measure indicating the direction and rate at which a function increases most steeply, used to adjust model parameters.
a machine learning optimization algorithm for minimizing the loss function, in which the model's parameters are adjusted so as to reduce the difference between the model's predicted values and the true data values.
an image feature representation that captures the distribution of object contours and edge directions, used to identify objects in an image.
a method of reducing the edges of a graph while preserving its core structure, in order to reduce computational cost and speed up the operation of graph-based machine learning algorithms.
false information generated by an AI and presented as fact.
a practical approach to problem-solving that uses approximate rules instead of exact calculations, when an exact solution would be too complex or costly.
a parameter set before the learning process that determines the model's learning process, architecture, data processing, or optimization settings (e.g., learning rate, number of layers, batch size).
the increasing of a system's capacity by adding more servers or nodes, allowing the system to handle larger data flows by distributing the load across multiple devices rather than upgrading a single server.
an AI agent that interacts with its environment through a physical (robot) or virtual (simulated) body.
a recurrent neural network architecture used for analyzing long data sequences, in which the relationship between elements far apart from one another is preserved.
smart systems that use sensors to detect and respond to people's needs without direct human intervention.
a hybrid machine learning method that uses a small quantity of labeled data and a large quantity of unlabeled data, in order to build accurate models while spending less time and resources on data labeling.
a function that converts multidimensional (2D, 3D) information into one-dimensional (1D) information.
a machine learning regularization technique that helps avoid overfitting, in which a portion of the artificial neurons is randomly switched off at each stage of training.
a reinforcement learning decision-making method in which the AI agent always chooses the action promising the greatest immediate reward, thereby maximizing short-term gain but limiting the discovery of new, potentially better solutions.
an activation function used to automatically normalize neuron outputs during training, helping to avoid vanishing or exploding gradient problems in artificial neural networks.
a problem-solving strategy in which all possible solution variants are systematically enumerated and checked.
the generation of predictions, decisions, or other content from new data using a trained model.
the final layer of an artificial neural network, which produces the final answer based on the information passed from the preceding layers.
1. a repeated step of a process; 2. a cycle of model development, training, and refinement, encompassing data preparation, training, and evaluation, aimed at optimizing the model's performance; 3. a single step of a machine learning cycle…
the representation of discrete data (e.g., words, images, or audio fragments) as numerical vectors in such a way that semantically similar objects are close in space, allowing an AI model to perform analysis of complex information…
a natural language processing process intended to identify and classify information in text into predefined categories (e.g., person names, organization names, dates).
the first layer of an artificial neural network, receiving the input data for further processing.
a decentralized machine learning method in which each device trains the model locally with its own data, and only the model weights are sent to a central server for aggregation, ensuring data privacy.
a process in which data obtained from several different types of sensors is integrated in order to obtain more accurate, reliable, and comprehensive information.
a mathematical function that computes the average error of a model's predictions across the entire dataset, where a higher cost means a greater number of model errors.
a neural network architecture in which neurons are grouped together, intended to represent not only objects but also their spatial relationships, orientations, and hierarchical structures.
an object detection model that sequentially applies progressively more complex classifiers, quickly rejecting irrelevant regions in order to more accurately analyze promising ones.
a table for evaluating the performance quality of a classification model, in which the model's predictions are compared with the true data values, marking correctly and incorrectly predicted positive and negative values.
1. a data-encoding device; 2. a part of a neural network that compresses high-dimensional input data (e.g., text, image) into low-dimensional form.
an interdisciplinary field of information technology that develops systems modeling human cognitive functions such as reasoning, learning, context awareness, or decision-making, capable of operating under uncertain conditions or making…
a field of AI that develops methods for extracting information from visual data, for example identifying objects or tracking movement, by modeling the human visual system.
the largest amount of information, measured in tokens, that an AI model can process at one time.
a mathematical operation in which a filter is slid across the input data with a fixed-size window, computing at each step the dot product of the corresponding elements.
a deep learning model that uses the convolution operation, specialized for the analysis of grid-structured data (e.g., images), used for automatic extraction of features from data.
the main part of a convolutional neural network, which by applying the convolution operation extracts different spatial features (e.g., edges, textures, shapes) from local regions of the input data and generates a feature map.
a loss function measuring the difference between the true probability distribution and the model's predicted probability distribution, most often used to optimize classification tasks during training.
a process in which the elements of one sequence are linked with the elements of another sequence, allowing the model to integrate information from both sources.
a process in which a model's parameters are converted from a higher-precision number format into a lower-precision format, thereby reducing the model's size and memory cost and speeding up its operation.
text processing in which words are transformed into their base dictionary form (the lemma).
a deep neural network architecture based on skip connections, which allow information to bypass certain layers, thereby facilitating the training of deep neural networks.
an iterative statistical optimization method consisting of two steps: the E-step, during which the expected values of hidden variables are computed, and the M-step, during which the model's parameters are optimized by maximizing the…
the dependence of a system's future state only on its current state, not on its past.
a mathematical decision-making model in which the future state depends only on the current state and action, not on the past.
the learning of computer systems from data without directly programming every action.
a computer system's ability to process and interpret data from the external world (vision, sound, sensor signals) at various levels --- from low-level tasks (e.g., edge detection) to high-level tasks (e.g., scene understanding).
the class with the fewest examples in the dataset.
a data normalization method in which values are transformed into a defined interval (usually [0, 1]) using the minimum and maximum values.
the process in which a machine learning model, using an optimization algorithm, adjusts its internal parameters according to the training data, in order to reduce the loss function and improve prediction accuracy.
a machine learning method in which a model is able to classify objects or perform tasks for which no examples were present in the training dataset, by relying on a semantic relationship between known and new categories or their…
a machine learning method in which a model is adapted to solve new tasks or recognize new data categories using only a small number of labeled training examples.
a machine learning method in which a model is adapted to solve new tasks or recognize new data categories using only a single labeled training example.
an algorithm that, during training, changes the learning rate according to a predefined schedule or strategy, in order to ensure better convergence and stability of the model.
an AI model capable of simultaneously processing and integrating different types of data (e.g., text, image, audio) in order to better understand the context of information and the relationships between different modalities.
a field of AI combining computer science and linguistics, intended to develop algorithms that allow computers to understand, interpret, analyze, and generate human language.
a generative AI method in which prompts are formulated by emphasizing undesired elements or characteristics of the generated result, in order to restrict the model's generation space and avoid unwanted artifacts.
a data balancing method in which, to balance the proportions of the majority and minority classes, a portion of examples is randomly removed from the majority class.
a machine learning model's inability to accurately identify patterns in data, due to insufficient model complexity or insufficient duration of the training cycle, resulting in poor performance on both training data and new data.
a machine learning method in which an AI model trained on unlabeled data independently discovers data structures, patterns, or groupings without predefined correct answers or labels.
a logical method for manipulating uncertain and imprecise information, in which the truth value of a statement can be any real number in the interval [0, 1] rather than only the discrete values "true" (1) or "false" (0).
a property of data in which one of the classes is represented with more examples than the others, thereby causing an AI model bias problem in which the majority class is favored and data from other classes is processed less well.
a problem in which neurons in artificial neural networks that use the rectified linear unit function become completely inactive (always returning the value 0) and can no longer learn, due to consistently negative input values.
a machine learning method in which a model is continuously updated with newly received data in real time, in order to maintain the relevance and accuracy of its predictions in a dynamic, changing environment.
a machine learning method in which a model sequentially acquires knowledge from different tasks or data streams over time, retaining previous knowledge and avoiding its forgetting.
a mathematical function measuring the difference between a model's prediction and the true value, indicating the model's error for a specific data example.
a modified rectified linear unit function that replaces negative input values with a small negative value, thereby avoiding the dying neurons problem.
a computer vision method used to identify and localize objects in an image by means of bounding boxes and the assignment of classes to them.
a computer vision technology: the automatic recognition of text in images and its conversion into editable digital text.
a computer vision method: the calculation and description, by means of vectors, of the movement of pixels between adjacent video frames.
a field of AI that aims to ensure that AI models' decisions are explained in a way understandable to humans, thereby increasing the transparency and reliability of systems.
the base part of a neural network architecture, intended to extract features from data, most often composed of pretrained convolutional layers.
a machine learning method in which the activations of each training batch are normalized according to the batch's mean and variance, thereby stabilizing and accelerating the training of neural networks.
a machine learning method that allows a model to learn sequentially from new data without repeating the entire training process and while preserving previously acquired knowledge.
a component of the long short-term memory model that regulates which information from the previous memory state is retained or discarded, by applying a sigmoid function.
a technology that overlays computer-generated elements (e.g., images, text, or 3D objects) onto a real-world environment in real time, using sensors and computer vision methods.
a neural network layer located between the input and output layers, performing intermediate data processing and feature extraction.
a visualization method in which each pixel or region of an image is marked with a value reflecting the model's prediction confidence at that location.
the execution of AI model tasks directly on end devices (e.g., phones, cameras, sensors) rather than on centralized cloud servers, thereby reducing latency, increasing data security, and lowering the need for internet access.
a machine learning method in which the parameters or features learned by a model trained on one task are reused to train a model for another, related task, reducing the amount of data and computational resources required.
a possible arrangement of a set of elements in a particular order, in which each element is used only once.
a quality metric for language models that measures the model's ability to predict the next word in a sequence.
the excessive adaptation of a model to the training data, resulting in poor performance when given new data.
a method for mitigating the machine learning problem arising from imbalanced data, in which the number of minority class examples is artificially increased.
a program that simulates a conversation with a human in text or voice form.
a machine learning process in which the features most relevant to the accuracy of the model's predictions are selected from the available set of features, reducing dimensionality and the risk of overfitting.
a multidimensional space in which each dimension represents one feature of the data.
the systematic construction and transformation of raw data representations into informative features, in order to improve machine learning models' ability to approximate the target relationship.
a machine learning process in which informative features suitable for model training are automatically extracted from raw data, reducing the dimensionality of the data while retaining important information.
an object detection architecture designed to analyze an image at different scales simultaneously, allowing the accurate detection of objects of varying sizes.
the output of a convolutional layer obtained after applying a filter, showing in the input data (e.g., an image) the location of specific detected features -- edges, texture, shapes.
a prompting method in which the desired elements or characteristics of the generated result are emphasized in the prompts.
a connection in an artificial neural network that allows information to bypass certain intermediate layers.
a reinforcement learning algorithm comprising two neural networks: an actor, which makes decisions, and a critic, which evaluates the advantage of those actions relative to the average expected return.
a machine learning method in which a model is trained from labeled data examples, and its parameters are continuously adjusted by an optimization algorithm to reduce the difference between predictions and true labels.
a field of data analysis combining statistical methods, machine learning, and data mining, intended to predict future events or trends based on historical data.
a set of conventions and procedures that allows separate programs to interact with each other or with the operating system.
the reduction of model overfitting by adding constraints or penalties.
an activation function that returns positive values unchanged and replaces negative values with zero.
a neural network with feedback connections that processes sequential data (text, time series) while retaining information about previous elements of the sequence.
the automatic detection and classification of patterns, structures, or features in data, using statistical or machine learning methods.
a learning method in which a model learns from unlabeled data by automatically generating training tasks from the structure of the data itself.
the ability of systems to coordinate their own activities without human intervention.
search that processes information by meaning, sense, and context, rather than only by keywords.
the identification and labeling, with pixel-level precision, of all objects in an image, performed by computer vision, grouping objects of the same class together.
an AI system capable of performing only specific, predefined tasks.
an activation function calculated as the product of the input value and the sigmoid function, ensuring more stable model training by applying a smooth and differentiable transition between negative and positive values.
an AI approach based on explicit symbolic knowledge representations, logical inference, and rules.
the splitting of text into individual characters (letters, digits, punctuation marks).
algorithmically generated data that imitates the statistical properties and structure of real data, used for training and testing models, especially when real data is scarce or hard to access.
the multiplication of two vectors of equal length, in which the products of the elements are summed to yield a single numerical value (a scalar).
the dividing boundary by which classes are separated according to their features.
a phenomenon in which the gradient value becomes too large during the training of artificial neural networks, causing the training to become unstable.
a deep learning method that applies the visual style of one image to the content of another image.
machine learning in which an agent learns to perform actions in an environment by trial and error, in order to maximize the cumulative reward over time.
machine learning in which a model's policy is optimized according to human evaluations, using a reward model that predicts how a human would rate the content.
a machine learning ensemble training method aimed at reducing bias, in which models are trained one after another, each correcting the errors of the previous one.
a hypothetical artificial intelligence whose cognitive abilities would surpass human intelligence in most domains.
a layer of an artificial neural network used to reduce the dimensionality of the data by aggregating the values of neighboring elements while retaining the most important features.
a loss function used to address the class imbalance problem, intended to focus greater learning attention on more difficult information.
the conversion of text into spoken language using generative AI.
the creation of videos from text descriptions using generative AI.
the creation of images from text descriptions using generative AI.
the passing of data through artificial neural layers from the input to the output layer, in order to generate a particular result.
an artificial neural network in which data flows in one direction, without entering a cycle, from the input layer, through the intermediate layers, to the output layer.
accurate, verified data used to train, test, and validate AI models.
a reinforcement learning decision-making method in which the AI agent performs new, untried actions in order to discover potentially better but as yet unknown long-term rewards.
a more efficient model architecture used to process sequences by applying the attention mechanism, processing the entire data sequence at once.
the removal of unwanted distortions from data (images, audio, signals), in order to improve its quality.
an instruction provided to an AI model.
the presentation of a prompt to a model without any examples, in which case the model relies only on its prior knowledge and the instruction.
a tuning method in which additional prompt vectors are trained in order to optimize the model's responses for specific tasks, without requiring full retraining of the model.
the process of creating, refining, and optimizing prompts in order to obtain the most accurate responses from AI models.
the refinement of prompt content in order to improve the quality of AI models' responses.
the presentation of a task to an AI model with several examples, increasing the accuracy of its responses without additional training.
the presentation of a task with a single example, helping the model understand a new task.
a computer vision method that automatically identifies objects, patterns, and other features in photographs or videos.
a computer vision method that partitions an image into meaningful separate regions for more precise analysis and processing.
a numerical value expressing the similarity of two or more images.
a recurrent neural network element that uses update and reset gates to control the flow of information in sequences, requiring fewer computational resources than the long short-term memory model.
a recurrent neural network composed of gated elements that use a gating mechanism to capture long-range sequence dependencies and control the flow of information.
a language used to describe which actions a system can perform and what immediate consequences those actions produce.
a machine learning method in which an agent's knowledge of the prerequisites and consequences of the actions available in its environment is created and modified.
a database composed of data in the form of vectors (sequences of numbers), which allows efficient searching for similar content by semantic similarity, including unstructured data such as photos, video, or audio recordings.
the increasing of a server's or system's performance by adding more resources (processor, memory, storage) to the existing server.
an attention mechanism in which each element of a sequence is linked with all other elements of the sequence, taking into account their mutual importance and relationships.
a model evaluation calculated by averaging the precision values across solutions to tasks of varying difficulty.
the restricted portion of the input with which an individual neuron of a convolutional neural network interacts, allowing the network to extract local spatial features.
a basic unit of information, a unit of text (a word, part of a word, or a character) into which an AI model breaks down text for processing.
the splitting of text into smaller units (tokens) that an AI model can process.
the maximum number of tokens that a model can process during an interaction, including input, context, and output.
a training data annotation indicating the correct answer or category.
an AI system that uses expert knowledge and a logical inference mechanism to solve problems in a specific domain.