Concept Glossary: C¶

This page lists English glossary entries for this letter. Entry bodies are assembled from term source files.

causal inference¶

• Meaning: Causal inference asks what changes the result if we change it? and tries to separate observed association from real causal effect.
• Why it matters: Feature importance tells us what a predictive model used a lot, but it does not prove that changing that feature will change the outcome.
• Related concepts: feature, correlation, intervention, interpretation boundary
• Core Section: P4-15.2
• Appears in: P4-15.2

chain rule¶

• Meaning: The chain rule is a differentiation rule for connected functions. When a calculation flows through several steps, it tells how a change in an earlier value affects a later result through the intermediate steps.
• Why it matters: Backpropagation uses the chain rule to compute gradients across many layers. It helps separate the mathematical rule from the procedure that applies it through a neural network.
• Related concepts: backpropagation, gradient, computation graph
• Core Section: P2-4.6
• Appears in: P5-5.1, P5-5.2

chunk¶

• Meaning: A chunk is a smaller passage or text unit cut from a longer document for search, retrieval, or comparison. It repackages a document into units that can be matched more directly with a question.
• Why it matters: RAG quality depends not only on the embedding model but also on how documents are chunked. Chunks that are too small can lose context, while chunks that are too large can mix in noise and make relevant passages harder to retrieve.
• Related concepts: embedding, document, retrieval-augmented generation, RAG
• Core Section: P1-13.1
• Appears in: P1-13.2, P1-13.3, P1-14.2, P6-2.3, P6-2.4, P6-2.5, P6-13.1, P6-13.2

classification¶

• Meaning: Classification is a problem type where an input case is assigned to one of a predefined set of categories or labels. It chooses from known options instead of generating a new free-form output.
• Why it matters: Many practical AI tasks, such as spam detection, anomaly alerts, and approval decisions, become classification questions. The concept also separates categorical prediction from regression, clustering, generation, and threshold-based operating decisions.
• Related concepts: clustering, prediction, supervised learning label, regression, threshold
• Core Section: P1-8.1
• Appears in: P1-1.2, P3-9.1, P5-15.1, P4-2.1, P4-3.2, P4-6.1, P4-6.2, P4-6.4, P4-8.1, P4-8.3, P4-9.1, P4-9.2, P4-11.1

cluster¶

• Meaning: A cluster is a group of data points treated as similar under the current distance, similarity, or density rule. It is a discovered group, not a human-defined answer category.
• Why it matters: Cluster outputs are easy to overread as real-world categories. This concept keeps the reader from treating cluster 0 or cluster 1 as meaningful names before the groups have been inspected and interpreted.
• Related concepts: clustering, cluster label, similarity
• Core Section: P4-17.1
• Appears In:

clustering¶

• Meaning: Clustering is an unsupervised problem type that finds groups of similar items in data without predefined answer labels. It asks how the data seems to separate by its own patterns.
• Why it matters: Clustering shows that learning can be about discovering structure rather than matching known answers. Clustering methods may group data by centers, dense regions, or gradually merged neighborhoods, but the resulting group number still requires human interpretation because it does not by itself explain the group meaning.
• Related concepts: classification, prediction, dataset, cluster
• Core Section: P4-17.1
• Appears in: P1-1.2, P4-index, P1-8.2, P4-2.2, P4-3.2, P4-6.2, P4-6.4, P4-8.1, P4-17.2, P4-18.2

CNN, convolutional neural network¶

• Meaning: A convolutional neural network is a neural-network structure that repeatedly reads local visual patterns and combines them into higher-level representations. Early layers can detect simple edges or textures, while later layers combine larger parts and object-level cues.
• Why it matters: CNNs are a central example of learned hierarchical representation in image recognition. They explain why local pattern reading, convolution, and pooling became powerful alternatives to treating an image as one flat fully connected input.
• Related concepts: image recognition, learned representation, deep learning
• Core Section: P1-9.1
• Appears in: P5-11.1, P5-11.2, P5-11.3

comparability¶

• Meaning: Comparability is the degree to which two or more values, samples, models, or conditions can be interpreted on the same basis.
• Why it matters: If row meaning, sample unit, measurement unit, or evaluation procedure differs, placing numbers side by side does not make the comparison fair. Checking comparability first helps distinguish real differences from results that mix incompatible standards.
• Related concepts: comparison result, data quality check, sample unit
• Core Section: P3-2.3

cross-validation¶

• Meaning: Cross-validation is a method that splits the available data multiple times and evaluates models or settings on different validation portions. It is often used when the dataset is small or one split may be too dependent on chance.
• Why it matters: Choosing a model from one validation score can be pulled around by the accident of a particular split. Cross-validation makes the comparison more cautious by repeating validation across several splits, but it does not magically remove the need for a final test check.
• Related concepts: validation data, test data, model selection, generalization
• Core Section: P4-4.2
• Appears in: P4-4.2, P4-5.2, P4-8.1, P4-9.1, P4-9.2, P4-9.3

cross-entropy¶

• Meaning: Cross-entropy is a probability-based loss that becomes larger when the model assigns a low probability to the correct answer. It is often used in classification and next-token prediction.
• Why it matters: It reads not only whether the top answer was correct, but also how much probability the model gave to the correct candidate. This connects classification loss, softmax output, and LLM next-token loss.
• Related concepts: loss function, softmax, log loss, next-token prediction
• Core Section: P5-4.2
• Appears in: P5-4.1

computation graph¶

• Meaning: A computation graph represents operations and value dependencies as connected nodes, making forward and backward calculation paths visible. It breaks a large formula into smaller calculation blocks and their connections.
• Why it matters: Computation graphs make backpropagation and automatic differentiation easier to understand. They show which intermediate values depend on which earlier operations and how gradient signals can flow backward through the graph.
• Related concepts: backpropagation, gradient, automatic differentiation
• Core Section: P5-5.2
• Appears in: P5-6.1

computational limit¶

• Meaning: A computational limit is the point where the candidate space grows too large to inspect every case within realistic time and resources. The issue is not merely that the computer is slow; the search space itself becomes too large.
• Why it matters: Computational limits explain why checking every candidate can be possible in theory but unusable in practice. They motivate heuristics, pruning, priority strategies, and approximate solutions in real systems.
• Related concepts: search, search space, heuristic
• Core Section: P1-7.1
• Appears in: P1-index, P1-6.1, P1-7.2, P1-7.4, P1-8.1

confabulation¶

• Meaning: Confabulation is a generation failure where a model confidently constructs content that sounds factual but lacks support. It is fluent fabrication rather than a simple typo or broken sentence.
• Why it matters: The concept helps distinguish natural-sounding text from grounded text. It explains why source checking and human review remain necessary even when an answer is smooth and even when retrieval has provided some candidate evidence.
• Related concepts: hallucination, supporting evidence, factuality, human oversight
• Core Section: P1-10.3
• Appears in: P1-12.3, P1-13.2, P7-5.3

confidential information¶

• Meaning: Confidential information is internal information that may not be legal personal data but can harm an organization, customer, contract, or operation if exposed. Draft contracts, unreleased code, customer lists, secrets, and launch plans are examples.
• Why it matters: Generative AI inputs and logs can copy or expose organizational information as well as personal data. The concept prevents the mistake of treating non-personal internal material as safe to paste into any tool.
• Related concepts: sensitive information, security, privacy, provenance, permission
• Core Section: P1-15.3
• Appears in: P1-14.5, P1-15.2, P1-16.2

confusion matrix¶

• Meaning: A confusion matrix is a table that counts classification results by actual label and predicted label. It shows error directions that a single accuracy number can hide, such as missed positive cases and incorrectly flagged positive cases.
• Why it matters: Two models with similar accuracy can fail in different cells, and those failures can carry different operational costs. A confusion matrix makes readers inspect the error structure before treating a score as sufficient.
• Related concepts: classification, metric, error, false positive, false negative
• First explained in: P4-6.1
• Appears in: P4-index, P4-6.1, P4-6.2, P4-8.2

context window¶

• Meaning: A context window is the maximum token range a model can keep available during one input-output computation. It is the size of the working memory for the current response, not unlimited memory.
• Why it matters: Long documents, long conversations, and retrieved passages must be selected or summarized to fit the context window. A larger window helps capacity but does not remove the need to choose what matters most.
• Related concepts: token, tokenization, retrieval-augmented generation, RAG
• Core Section: P6-4.2
• Appears in: P6-2.1, P6-4.1

contextual representation¶

• Meaning: A contextual representation is an internal vector representation that changes with surrounding context even for the same word or token. The same surface form can receive different representations in different sentences.
• Why it matters: This concept marks the move beyond fixed word meanings. It helps readers understand why encoder-centered language models and LLMs can represent a word according to its current role in a sentence.
• Related concepts: embedding, encoder, language model
• Core Section: P1-11.3
• Appears in: P6-6.1

convergence¶

• Meaning: Convergence is the process where values or a sequence get closer to a particular value or stable state over repeated steps. It does not require reaching the exact value immediately; the key is whether changes shrink over time.
• Why it matters: Convergence helps readers interpret iterative computation and training curves. It separates movement toward a stable point from divergence, oscillation, or unstable learning behavior.
• Related concepts: limit, rate of change, optimization
• Core Section: P2-2.3
• Appears in: P2-6.3, P4-19.5

conversational LLM¶

• Meaning: A conversational LLM is the user-facing form created when instruction following, dialogue formatting, safety tuning, and interface layers are added on top of an LLM. It is a model-plus-product layer, not only the base model.
• Why it matters: The concept separates a language model from a chatbot experience. System prompts, safety rules, tools, and UI flow can change how the same base model feels and behaves to users.
• Related concepts: language model, pretraining, instruction tuning
• Core Section: P6-5.2

convolution¶

• Meaning: Convolution is an operation that moves a small filter across an input and computes local pattern responses at each position. It repeatedly asks whether a particular pattern appears in each local window.
• Why it matters: Convolution is the core operation that lets CNNs read local visual patterns without mixing the whole image at once. It helps explain feature maps and why the same detector can be applied across many image locations.
• Related concepts: CNN, convolutional neural network, feature, image recognition
• Core Section: P5-11.2

copyright¶

• Meaning: Copyright is the legal right that protects creative expression such as text, images, charts, code, and other concrete works. It protects expression rather than the abstract idea alone.
• Why it matters: Copyright matters when using books, articles, images, code, or diagrams in AI inputs and public drafts. Public availability does not automatically mean the expression can be reused freely.
• Related concepts: quotation, license, training data
• Core Section: P1-15.2
• Appears in: P1-10.3

corpus¶

• Meaning: A corpus is a collection of text gathered for training, analyzing, or evaluating language models and embeddings. It represents selected language use, not just a random pile of documents.
• Why it matters: Language models learn patterns from the distribution of text they see. Corpus domain, time period, style, cleaning, bias, and coverage strongly shape model strengths and limitations.
• Related concepts: language model, data, embedding
• Core Section: P1-11.1
• Appears in: P1-11.3, P6-2.5, P6-20.1

cosine similarity¶

• Meaning: Cosine similarity measures how similarly two vectors point in direction. It focuses on the angle between vectors more than their absolute lengths.
• Why it matters: In text embedding search, direction can matter more than magnitude. Cosine similarity explains how semantic similarity can be read as closeness of direction, while also reminding readers that magnitude-sensitive tasks may need other measures.
• Related concepts: similarity, distance, similarity search
• Core Section: P1-13.2
• Appears in: P1-13.1, P2-3.2, P2-3.4

centroid¶

• Meaning: A centroid is the representative center of a cluster, usually computed as the mean position of the points assigned to that cluster.
• Why it matters: A centroid is a useful summary, but it is not the same as every point in the group. Readers need this distinction to interpret center-based clustering results without forgetting the spread and shape of the actual data.
• Related concepts: cluster, clustering, distance
• Core Section: P4-17.1
• Appears In:

credential¶

• Meaning: A credential is a proof value such as a password, API key, token, or cookie used to verify identity, access, or request validity. It can directly grant access to accounts, sessions, or external services.
• Why it matters: If credentials appear in AI inputs, logs, or test outputs, exposure can lead to account compromise or service misuse. This concept connects masking, least privilege, and input restrictions to concrete operational risk.
• Related concepts: sensitive information, confidential information, security, provenance, permission
• Core Section: P1-15.3
• Appears in: P1-14.5, P1-16.2

contrastive learning¶

• Meaning: Contrastive learning uses pairs that should become close and pairs that should become far apart to learn the placement of a representation space.
• Why it matters: Embedding quality depends not only on search speed, but on what the model learned to place close or far away. Contrastive learning gives the basic sense that same-intent or same-problem expressions should become close, while different-intent expressions should separate.
• Related concepts: positive pair, negative pair, embedding, representation learning
• Core Section: P6-3.3
• Appears in: P6-3.3

correlation coefficient¶

• Meaning: A standardized number that summarizes the direction and strength of a linear relationship between two variables.
• Why it matters: It helps readers distinguish visible co-movement from causal explanation and compare relationships across differently scaled variables.
• Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
• Core Section: P2-5.5

context dependency¶

• Meaning: The property that the meaning or output of an item changes depending on the surrounding input, earlier text, or task conditions.
• Why it matters: It is central to reading LLM behavior, because the same token or sentence can behave differently under different prompts and contexts.
• Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
• Core Section: P1-10.2

composite function¶

• Meaning: A function made by feeding the output of one function into another function.
• Why it matters: It matters because neural networks, loss calculations, and the chain rule are easier to read when layered calculations are seen as composed functions.
• Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
• Core Section: P2-4.6

confidence interval¶

• Meaning: A range computed from sample data that expresses the uncertainty around an estimated value under a stated procedure.
• Why it matters: It prevents a single estimate from being read as exact and helps connect sample size, uncertainty, and interpretation strength.
• Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
• Core Section: P2-5.5

covariance¶

• Meaning: A measure of how two variables tend to move together around their means.
• Why it matters: It opens the first reading of relationship between variables, while also showing why scale-sensitive measures often need standardization.
• Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
• Core Section: P2-5.5
• Appears in: P4-18.1

CPU offloading¶

• Meaning: CPU offloading is a memory management strategy that keeps model parts or weights outside GPU VRAM when they are not immediately needed, then moves them to the GPU at execution time.
• Why it matters: It helps separate the question of whether a large model can run on limited GPU memory from the question of whether the generated output is good. CPU offloading can reduce out-of-memory failures, but it does not directly improve prompt following, pose control, or output quality.
• Related concepts: computational limit, tensor, inference, open-weight model
• Core Section: P6-21.2
• Appears in: P7-5.1, P7-5.3, P7-5.4, P7-5.11