Читать книгу Глоссариум по искусственному интеллекту: 2500 терминов. Том 2 - Александр Николаевич Чашин, Инна Евгеньевна Скворцова, Александр Николаевич Афонин - Страница 5

CAFFE is short for Convolutional Architecture for Fast Feature Embedding which is an open-source deep learning framework de- veloped in Berkeley AI Research. It supports many different deep learning architectures and GPU-based acceleration computation kernels188,189.

Calibration layer is a post-prediction adjustment, typically to account for prediction bias. The adjusted predictions and probabilities should match the distribution of an observed set of labels190.

Candidate generation — the initial set of recommendations chosen by a recommendation system191.

Candidate sampling is a training-time optimization in which a probability is calculated for all the positive labels, using, for example, softmax, but only for a random sample of negative labels. For example, if we have an example labeled beagle and dog candidate sampling computes the predicted probabilities and corresponding loss terms for the beagle and dog class outputs in addition to a random subset of the remaining classes (cat, lollipop, fence). The idea is that the negative classes can learn from less frequent negative reinforcement as long as positive classes always get proper positive reinforcement, and this is indeed observed empirically. The motivation for candidate sampling is a computational efficiency win from not computing predictions for all negatives192.

Canonical Formats in information technology, canonicalization is the process of making something conform] with some specification… and is in an approved format. Canonicalization may sometimes mean generating canonical data from noncanonical data. Canonical formats are widely supported and considered to be optimal for long-term preservation193.

Capsule neural network (CapsNet) is a machine learning system that is a type of artificial neural network (ANN) that can be used to better model hierarchical relationships. The approach is an attempt to more closely mimic biological neural organization194,195.

Case-Based Reasoning (CBR) is a way to solve a new problem by using solutions to similar problems. It has been formalized to a process consisting of case retrieve, solution reuse, solution revise, and case retention196.

Categorical data — features having a discrete set of possible values. For example, consider a categorical feature named house style, which has a discrete set of three possible values: Tudor, ranch, colonial. By representing house style as categorical data, the model can learn the separate impacts of Tudor, ranch, and colonial on house price. Sometimes, values in the discrete set are mutually exclusive, and only one value can be applied to a given example. For example, a car maker categorical feature would probably permit only a single value (Toyota) per example. Other times, more than one value may be applicable. A single car could be painted more than one different color, so a car color categorical feature would likely permit a single example to have multiple values (for example, red and white). Categorical features are sometimes called discrete features. Contrast with numerical data197.

Center for Technological Competence is an organization that owns the results, tools for conducting fundamental research and platform solutions available to market participants to create applied solutions (products) on their basis. The Technology Competence Center can be a separate organization or be part of an application technology holding company198.

Central Processing Unit (CPU) is a von Neumann cyclic processor designed to execute complex computer programs199.

Centralized control is a process in which control signals are generated in a single control center and transmitted from it to numerous control objects200.

Centroid – the center of a cluster as determined by a k-means or k-median algorithm. For instance, if k is 3, then the k-means or k-median algorithm finds 3 centroids201.

Centroid-based clustering is a category of clustering algorithms that organizes data into nonhierarchical clusters. k-means is the most widely used centroid-based clustering algorithm. Contrast with hierarchical clustering algorithms202.

Character format is any file format in which information is encoded as characters using only a standard character-encoding scheme. A file written in «character format» contains only those bytes that are prescribed in the encoding scheme as corresponding to the characters in the scheme (e.g., alphabetic and numeric characters, punctuation marks, and spaces)203.

Сhatbot is a software application designed to simulate human conversation with users via text or speech. Also referred to as virtual agents, interactive agents, digital assistants, or conversational AI, chatbots are often integrated into applications, websites, or messaging platforms to provide support to users without the use of live human agents. Chatbots originally started out by offering users simple menus of choices, and then evolved to react to particular keywords. «But humans are very inventive in their use of language,» says Forrester’s McKeon-White. Someone looking for a password reset might say they’ve forgotten their access code, or are having problems getting into their account. «There are a lot of different ways to say the same thing,» he says. This is where AI comes in. Natural language processing is a subset of machine learning that enables a system to understand the meaning of written or even spoken language, even where there is a lot of variation in the phrasing. To succeed, a chatbot that relies on AI or machine learning needs first to be trained using a data set. In general, the bigger the training data set, and the narrower the domain, the more accurate and helpful a chatbot will be204.

Checkpoint — data that captures the state of the variables of a model at a particular time. Checkpoints enable exporting model weights, as well as performing training across multiple sessions. Checkpoints also enable training to continue past errors (for example, job preemption). Note that the graph itself is not included in a checkpoint205.

Chip is an electronic microcircuit of arbitrary complexity, made on a semiconductor substrate and placed in a non-separable case or without it, if included in the micro assembly206,207.

Class — one of a set of enumerated target values for a label. For example, in a binary classification model that detects spam, the two classes are spam and not spam. In a multi-class classification model that identifies dog breeds, the classes would be poodle, beagle, pug, and so on208.

Classification model is a type of machine learning model for distinguishing among two or more discrete classes. For example, a natural language processing classification model could determine whether an input sentence was in French, Spanish, or Italian209.

Classification threshold is a scalar-value criterion that is applied to a model’s predicted score in order to separate the positive class from the negative class. Used when mapping logistic regression results to binary classification210.

Classification. Classification problems use an algorithm to accurately assign test data into specific categories, such as separating apples from oranges. Or, in the real world, supervised learning algorithms can be used to classify spam in a separate folder from your inbox. Linear classifiers, support vector machines, decision trees and random forest are all common types of classification algorithms211.

Сloud robotics is a field of robotics that attempts to invoke cloud technologies such as cloud computing, cloud storage, and other Internet technologies centred on the benefits of converged infrastructure and shared services for robotics. When connected to the cloud, robots can benefit from the powerful computation, storage, and communication resources of modern data center in the cloud, which can process and share information from various robots or agent (other machines, smart objects, humans, etc.). Humans can also delegate tasks to robots remotely through networks. Cloud computing technologies enable robot systems to be endowed with powerful capability whilst reducing costs through cloud technologies. Thus, it is possible to build lightweight, low cost, smarter robots have intelligent «brain» in the cloud. The «brain» consists of data center, knowledge base, task planners, deep learning, information processing, environment models, communication support, etc.212.

Clinical Decision Support (CDS) is a health information technology system that is designed to provide physicians and other health professionals with clinical decision support, that is, assistance with clinical decision- making tasks213.

Clipping is a technique for handling outliers. Specifically, reducing feature values that are greater than a set maximum value down to that maximum value. Also, increasing feature values that are less than a specific minimum value up to that minimum value. For example, suppose that only a few feature values fall outside the range 40—60. In this case, you could do the following: Clip all values over 60 to be exactly 60. Clip all values under 40 to be exactly 40. In addition to bringing input values within a designated range, clipping can also used to force gradient values within a designated range during training214.

Closed dictionary in speech recognition systems, a dictionary with a limited number of words, to which the recognition system is configured and which cannot be replenished by the user215.

Cloud computing is an information technology model for providing ubiquitous and convenient access using the Internet to a common set of configurable computing resources («cloud»), data storage devices, applications and services that can be quickly provided and released from the load with minimal operating costs or with little or no involvement of the provider216.

Cloud is a general metaphor that is used to refer to the Internet. Initially, the Internet was seen as a distributed network and then with the invention of the World Wide Web as a tangle of interlinked media. As the Internet continued to grow in both size and the range of activities it encompassed, it came to be known as «the cloud.» The use of the word cloud may be an attempt to capture both the size and nebulous nature of the Internet217.

Cloud TPU is a specialized hardware accelerator designed to speed up machine learning workloads on Google Cloud Platform218.

Cluster analysis is a type of unsupervised learning used for exploratory data analysis to find hidden patterns or groupings in the data; clusters are modeled with a similarity measure defined by metrics such as Euclidean or probability distance.

Clustering is a data mining technique for grouping unlabeled data based on their similarities or differences. For example, K-means clustering algorithms assign similar data points into groups, where the K value represents the size of the grouping and granularity. This technique is helpful for market segmentation, image compression, etc219.

Co-adaptation is when neurons predict patterns in training data by relying almost exclusively on outputs of specific other neurons instead of relying on the network’s behavior as a whole. When the patterns that cause co-adaption are not present in validation data, then co-adaptation causes overfitting. Dropout regularization reduces co-adaptation because dropout ensures neurons cannot rely solely on specific other neurons220.

COBWEB is an incremental system for hierarchical conceptual clustering. COBWEB was invented by Professor Douglas H. Fisher, currently at Vanderbilt University. COBWEB incrementally organizes observations into a classification tree. Each node in a classification tree represents a class (concept) and is labeled by a probabilistic concept that summarizes the attribute-value distributions of objects classified under the node. This classification tree can be used to predict missing attributes or the class of a new object221.

Code is a one-to-one mapping of a finite ordered set of symbols belonging to some finite alphabet222.

Codec is a codec is the means by which sound and video files are compressed for storage and transmission purposes. There are various forms of compression: ’lossy’ and ’lossless’, but most codecs perform lossless compression because of the much larger data reduction ratios that occur with lossy compression. Most codecs are software, although in some areas codecs are hardware components of image and sound systems. Codecs are necessary for playback, since they uncompress or decompress the moving image and sound files and allow them to be rendered223.

Cognitive architecture – the Institute of Creative Technologies defines cognitive architecture as: «hypothesis about the fixed structures that provide a mind, whether in natural or artificial systems, and how they work together – in conjunction with knowledge and skills embodied within the architecture – to yield intelligent behavior in a diversity of complex environments»224.

Cognitive computing is used to refer to the systems that simulate the human brain to help with the decision- making. It uses self-learning algorithms that perform tasks such as natural language processing, image analysis, reasoning, and human—computer interaction. Examples of cognitive systems are IBM’s Watson and Google DeepMind225.

Cognitive Maps are structured representations of decision depicted in graphical format (variations of cognitive maps are cause maps, influence diagrams, or belief nets). Basic cognitive maps include nodes connected by arcs, where the nodes represent constructs (or states) and the arcs represent relationships. Cognitive maps have been used to understand decision situations, to analyze complex cause-effect representations and to support communication226.

Cognitive science – the interdisciplinary scientific study of the mind and its processes227.

Cohort is a sample in study (conducted to evaluate a machine learning algorithm, for example) where it is followed prospectively or retrospectively and subsequent status evaluations with respect to a disease or outcome are conducted to determine which initial participants’ exposure characteristics (risk factors) are associated with it.

Cold-Start is a potential issue arising from the fact that a system cannot infer anything for users or items for which it has not gathered a sufficient amount of information yet228.

Collaborative filtering – making predictions about the interests of one user based on the interests of many other users. Collaborative filtering is often used in recommendation systems229.

Combinatorial optimization in operations research, applied mathematics and theoretical computer science, combinatorial optimization is a topic that consists of finding an optimal object from a finite set of objects230.

Committee machine is a type of artificial neural network using a divide and conquer strategy in which the responses of multiple neural networks (experts) are combined into a single response. The combined response of the committee machine is supposed to be superior to those of its constituent experts. Compare ensembles of classifiers231.

Commoditization is the process of transforming a product from an elite to a generally available (comparatively cheap commodity of mass consumption)232.

Common Data Element (CDE) is a tool to support data management for clinical research233.

Commonsense reasoning is a branch of artificial intelligence concerned with simulating the human ability to make presumptions about the type and essence of ordinary situations they encounter every day234.

Compiler is a program that translates text written in a programming language into a set of machine codes. AI framework compilers collect the computational data of the frameworks and try to optimize the code of each of them, regardless of the hardware of the accelerator. The compiler contains programs and blocks with which the framework performs several tasks. The computer memory resource allocator, for example, allocates power individually for each accelerator235.

Composite AI is a combined application of various artificial intelligence methods (deep machine learning, computer vision, natural language processing, contextual analysis, knowledge graphs, data visualization, forecasting methods, etc.) to increase the efficiency of model training in order to achieve a synergistic effect from their use and the best results of the work of artificial intelligence systems. One of the ideas that is laid down in the creation of composite artificial intelligence is to obtain a sane artificial intelligence that will be able to understand the essence of the problems and solve a wide range of problems, offering optimal solutions.236,237,238.

Compression is a method of reducing the size of computer files. There are several compression programs available, such as gzip and WinZip239.

Computation is any type of arithmetic or non-arithmetic calculation that follows a well-defined model (e.g., an algorithm)240.

Computational chemistry is a discipline using mathematical methods for the calculation of molecular properties or for the simulation of molecular behaviour. It also includes, e.g., synthesis planning, database searching, combinatorial library manipulation.241,242,243.

Computational complexity theory – focuses on classifying computational problems according to their inherent difficulty, and relating these classes to each other. A computational problem is a task solved by a computer. A computation problem is solvable by mechanical application of mathematical steps, such as an algorithm244.

Computational creativity (also artificial creativity, mechanical creativity, creative computing, or creative computation) is a multidisciplinary endeavour that includes the fields of artificial intelligence, cognitive psychology, philosophy, and the arts245.

Computational cybernetics is the integration of cybernetics and computational intelligence techniques246.

Computational efficiency of an agent or a trained model is the amount of computational resources required by the agent to solve a problem at the inference stage247.

Computational efficiency of an intelligent system is the amount of computing resources required to train an intelligent system with a certain level of performance on a given volume of tasks248.

Computational Graphics Processing Unit (computational GPU, cGPU) – graphic processor-computer, GPU-computer, multi-core GPU used in hybrid supercomputers to perform parallel mathematical calculations; for example, one of the first GPUs in this category contains more than 3 billion transistors – 512 CUDA cores and up to 6 GB of memory249.

Computational humor is a branch of computational linguistics and artificial intelligence which uses computers in humor research250.

Computational intelligence (CI) usually refers to the ability of a computer to learn a specific task from data or experimental observation251.

Computational learning theory (COLT) in computer science, is a subfield of artificial intelligence devoted to studying the design and analysis of machine learning algorithms252.

Computational linguistics is an interdisciplinary field concerned with the statistical or rule-based modeling of natural language from a computational perspective, as well as the study of appropriate computational approaches to linguistic questions253.

Computational mathematics is the mathematical research in areas of science where computing plays an essential role254.

Computational neuroscience (also known as theoretical neuroscience or mathematical neuroscience) is a branch of neuroscience which employs mathematical models, theoretical analysis and abstractions of the brain to understand the principles that govern the development, structure, physiology, and cognitive abilities of the nervous system255,256.

Computational number theory (also algorithmic number theory) – the study of algorithms for performing number theoretic computations257,,258.

Computational problem in theoretical computer science is a mathematical object representing a collection of questions that computers might be able to solve259.

Computational statistics (or statistical computing) is the application of computer science and software engineering principles to solving scientific problems. It involves the use of computing hardware, networking, algorithms, programming, databases and other domain-specific knowledge to design simulations of physical phenomena to run on computers. Computational science crosses disciplines and can even involve the humanities260,261.

Computer engineering — technologies for digital modeling and design of objects and production processes throughout the life cycle262.

Computer incident is a fact of violation and (or) cessation of the operation of a critical information infrastructure object, a telecommunication network used to organize the interaction of such objects, and (or) a violation of the security of information processed by such an object, including as a result of a computer attack263.

Computer science – the theory, experimentation, and engineering that form the basis for the design and use of computers. It involves the study of algorithms that process, store, and communicate digital information. A computer scientist specializes in the theory of computation and the design of computational systems. Computer science is the study of computation, automation, and information. Computer science spans theoretical disciplines (such as algorithms, theory of computation, and information theory) to practical disciplines (including the design and implementation of hardware and software). Computer science is generally considered an area of academic research and distinct from computer programming264.

Computer simulation is the process of mathematical modelling, performed on a computer, which is designed to predict the behaviour of, or the outcome of, a real-world or physical system. The reliability of some mathematical models can be determined by comparing their results to the real-world outcomes they aim to predict. Computer simulations have become a useful tool for the mathematical modeling of many natural systems in physics (computational physics), astrophysics, climatology, chemistry, biology and manufacturing, as well as human systems in economics, psychology, social science, health care and engineering265.

Computer vision (CV) is scientific discipline, field of technology and the direction of artificial intelligence (AI), which deals with computer processing, recognition, analysis and classification of dynamic images of reality. It is widely used in video surveillance systems, in robotics and in modern industry to improve product quality and production efficiency, comply with legal requirements, etc. In computer vision, the following areas are distinguished: face recognition (face recognition), image recognition (image recognition), augmented reality (augmented reality (AR) and optical character recognition (OCR). Synonyms – artificial vision, machine vision266.

Computer vision processing (CVP) is the processing of images (signals) in a computer vision system, in computer vision systems – about algorithms (computer vision processing algorithms), processors (computer vision processing unit, CVPU), convolutional neural networks (convolutional neural network), which are used for image processing and implementation of visual functions in robotics, real-time systems, smart video surveillance systems, etc.267.

Computer-Aided Detection/Diagnosis (CAD), uses computer programs to assist radiologists in the interpretation of medical images. CAD systems process digital images for typical appearances and highlight suspicious regions in order to support a decision taken by a professional268.

Computer-automated design (CAutoD) – design automation usually refers to electronic design automation, or Design Automation which is a Product Configurator. Extending Computer-Aided Design (CAD), automated design and computer-automated designare concerned with a broader range of applications, such as automotive engineering, civil engineering, composite material design, control engineering, dynamic system identification and optimization, financial systems, industrial equipment, mechatronic systems, steel construction, structural optimisation, and the invention of novel systems. More recently, traditional CAD simulation is seen to be transformed to CAutoD by biologically inspired machine learning, including heuristic search techniques such as evolutionary computation, and swarm intelligence algorithms269.

Computing modules are plug-in specialized computers designed to solve narrowly focused tasks, such as accelerating the work of artificial neural networks algorithms, computer vision, voice recognition, machine learning and other artificial intelligence methods, built on the basis of a neural processor – a specialized class of microprocessors and coprocessors (processor, memory, data transfer).

Computing system is a software and hardware complex intended for solving problems and processing data (including calculations) or several interconnected complexes that form a single infrastructure270.

Computing units are blocks that work like a filter that transforms packets according to certain rules. The instruction set of the calculator can be limited, which guarantees a simple internal structure and a sufficiently high speed of operation271.

Сoncept drift in predictive analytics and machine learning, the concept drift means that the statistical properties of the target variable, which the model is trying to predict, change over time in unforeseen ways. This causes problems because the predictions become less accurate as time passes272.

Сonnectionism is an approach in the fields of cognitive science, that hopes to explain mental phenomena using artificial neural networks273,274.

Сonsistent heuristic in the study of path-finding problems in artificial intelligence, a heuristic function is said to be consistent, or monotone, if its estimate is always less than or equal to the estimated distance from any neighboring vertex to the goal, plus the cost of reaching that neighbor275.

Сonstrained conditional model (CCM) is a machine learning and inference framework that augments the learning of conditional (probabilistic or discriminative) models with declarative сonstraints276.

Constraint logic programming is a form of constraint programming, in which logic programming is extended to include concepts from constraint satisfaction. A constraint logic program is a logic program that contains constraints in the body of clauses277.

Constraint programming is a programming paradigm wherein relations between variables are stated in the form of constraints. Constraints differ from the common primitives of imperative programming languages in that they do not specify a step or sequence of steps to execute, but rather the properties of a solution to be found278.

Constructed language (also conlang) is a language whose phonology, grammar, and vocabulary are consciously devised, instead of having developed naturally. Constructed languages may also be referred to as artificial, planned, or invented languages279.

Control theory in control systems engineering, is a subfield of mathematics that deals with the control of continuously operating dynamical systems in engineered processes and machines. The objective is to develop a control model for controlling such systems using a control action in an optimum manner without delay or overshoot and ensuring control stability280.

Convolutional neural network (CNN, or ConvNet) in deep learning, is a class of deep neural networks, most commonly applied to analyzing visual imagery. CNNs use a variation of multilayer perceptrons designed to require minimal preprocessing. They are also known as shift invariant or space invariant artificial neural networks (SIANN), based on their shared-weights architecture and translation invariance characteristics. Сonvolutional neural network is a class of artificial neural network most commonly used to analyze visual images. They are also known as Invariant or Spatial Invariant Artificial Neural Networks (SIANN) based on an architecture with a common weight of convolution kernels or filters that slide over input features and provide equivalent translation responses known as feature maps281.

Confidentiality of information is a mandatory requirement for a person who has access to certain information not to transfer such information to third parties without the consent of its owner282.

Confirmation Bias – the tendency to search for, interpret, favor, and recall information in a way that confirms one’s own beliefs or hypotheses while giving disproportionately less attention to information that contradicts it283.

Confusion matrix is a situational analysis table that summarizes the prediction results of a classification model in machine learning. The records in the dataset are summarized in a matrix according to the real category and the classification score made by the classification model284,285.

Consumer artificial intelligence is specialized artificial intelligence programs embedded in consumer devices and processes286.

Continuous feature is a floating-point feature with an infinite range of possible values. Contrast with discrete feature287,288.

Contributor is a human worker providing annotations on the Appen data annotation platform289.

Convenience sampling – using a dataset not gathered scientifically in order to run quick experiments. Later on, it’s essential to switch to a scientifically gathered dataset290.

Convergence – informally, often refers to a state reached during training in which training loss and validation loss change very little or not at all with each iteration after a certain number of iterations. In other words, a model reaches convergence when additional training on the current data will not improve the model. In deep learning, loss values sometimes stay constant or nearly so for many iterations before finally descending, temporarily producing a false sense of convergence. See also early stopping291,292.

Convex function is a function in which the region above the graph of the function is a convex set. The prototypical convex function is shaped something like the letter U. For example, the following are all convex functions:

By contrast, the following function is not convex. Notice how the region above the graph is not a convex set:

A strictly convex function has exactly one local minimum point, which is also the global minimum point. The classic U-shaped functions are strictly convex functions. However, some convex functions (for example, straight lines) are not U-shaped. A lot of the common loss functions, including the following, are convex functions: L2 loss; Log Loss; L1 regularization; L2 regularization. Many variations of gradient descent are guaranteed to find a point close to the minimum of a strictly convex function. Similarly, many variations of stochastic gradient descent have a high probability (though, not a guarantee) of finding a point close to the minimum of a strictly convex function. The sum of two convex functions (for example, L2 loss + L1 regularization) is a convex function. Deep models are never convex functions. Remarkably, algorithms designed for convex optimization tend to find reasonably good solutions on deep networks anyway, even though those solutions are not guaranteed to be a global minimum293,294.

Convex optimization – the process of using mathematical techniques such as gradient descent to find the minimum of a convex function. A great deal of research in machine learning has focused on formulating various problems as convex optimization problems and in solving those problems more efficiently. For complete details, see Boyd and Vandenberghe, Convex Optimization295.

Convex set is a subset of Euclidean space such that a line drawn between any two points in the subset remains completely within the subset.296.

Convolution — the process of filtering. A filter (or equivalently: a kernel or a template) is shifted over an input image. The pixels of the output image are the summed product of the values in the filter pixels and the corresponding values in the underlying image297.

Convolutional filter – one of the two actors in a convolutional operation. (The other actor is a slice of an input matrix). A convolutional filter is a matrix having the same rank as the input matrix, but a smaller shape298.

Convolutional layer is a layer of a deep neural network in which a convolutional filter passes along an input matrix299.

Convolutional neural network (CNN) is a type of neural network that identifies and interprets images300,301.

Convolutional operation – the following two-step mathematical operation: Element-wise multiplication of the convolutional filter and a slice of an input matrix. (The slice of the input matrix has the same rank and size as the convolutional filter); Summation of all the values in the resulting product matrix302.

Corelet programming environment (CPE) is a scalable environment that allows programmers to set the functional behavior of a neural network by adjusting its parameters and communication characteristics303.

Corpus of texts is a large dataset of written or spoken material that can be used to train a machine to perform linguistic tasks304.

Correlation analysis is a statistical data processing method that measures the strength of the relationship between two or more variables. Thus, it determines whether there is a connection between the phenomena and how strong the connection between these phenomena is305.

Correlation is a statistical relationship between two or more random variables306.

Cost – synonym for loss. A measure of how far a model’s predictions are from its label. Or, to put it more pessimistically, a measure of how bad a model is. To determine this value, the model must define a loss function. For example, linear regression models typically use the standard error for the loss function, while logistic regression models use the log loss307,308.

Co-training essentially amplifies independent signals into a stronger signal. For instance, consider a classification model that categorizes individual used cars as either Good or Bad. One set of predictive features might focus on aggregate characteristics such as the year, make, and model of the car; another set of predictive features might focus on the previous owner’s driving record and the car’s maintenance history. The seminal paper on co-training is Combining Labeled and Unlabeled Data with Co-Training by Blum and Mitchell309.

Counterfactual fairness is a fairness metric that checks whether a classifier produces the same result for one individual as it does for another individual who is identical to the first, except with respect to one or more sensitive attributes. Evaluating a classifier for counterfactual fairness is one method for surfacing potential sources of bias in a model. See «When Worlds Collide: Integrating Different Counterfactual Assumptions in Fairness» for a more detailed discussion of counterfactual fairness310.

Coverage bias – this bias means that the study sample is not representative and that the data set in the array has zero chance of being included in the sample311.

Crash blossom is a sentence or phrase with an ambiguous meaning. Crash blossoms present a significant problem in natural language understanding. For example, the headline Red Tape Holds Up Skyscraper is a crash blossom because an NLU model could interpret the headline literally or figuratively312.

Critic – synonym for Deep Q-Network313.

Critical information infrastructure – objects of critical information infrastructure, as well as telecommunication networks used to organize the interaction of such objects314.

Critical information infrastructure of the Russian Federation is a set of critical information infrastructure objects, as well as telecommunication networks used to organize the interaction of critical information infrastructure objects with each other315.

Cross-entropy is a generalization of Log Loss to multi-class classification problems. Cross-entropy quantifies the difference between two probability distributions. See also perplexity316.

Crossover (also recombination) in genetic algorithms and evolutionary computation, a genetic operator used to combine the genetic information of two parents to generate new offspring. It is one way to stochastically generate new solutions from an existing population, and analogous to the crossover that happens during sexual reproduction in biological organisms. Solutions can also be generated by cloning an existing solution, which is analogous to asexual reproduction. Newly generated solutions are typically mutated before being added to the population317.

Cross-Validation (k-fold Cross-Validation, Leave-p-out Cross-Validation) is a collection of processes designed to evaluate how the results of a predictive model will generalize to new data sets. k-fold Cross-Validation; Leave-p-out Cross-Validation318.

Cryogenic freezing (cryonics, human cryopreservation) is a technology of preserving in a state of deep cooling (using liquid nitrogen) the head or body of a person after his death with the intention to revive them in the future319.

Cyber-physical systems are intelligent networked systems with built-in sensors, processors and drives that are designed to interact with the physical environment and support the operation of computer information systems in real time320.