Concept Glossary: S¶

This page lists English glossary entries for this letter. Entry bodies are generated from per-term source files with includes.

safety¶

Meaning: AI system safety means limiting and reviewing AI outputs or automation so they do not create real harm for people, organizations, or procedures.
Why it matters: AI errors can move beyond wording problems into decisions, recommendations, tool execution, privacy exposure, or unmanaged automation. Safety keeps harmful execution paths, excessive permission, and missing review in view.
Related concepts: bias, accountability, human oversight, permission
Core Section: P1-15.1

sim-to-real gap¶

Meaning: The sim-to-real gap is the difference that appears when a policy trained or evaluated in simulation is moved into the real world, where noise, delay, friction, lighting, and rare cases can differ. Deliberately varying simulation conditions is one sub-strategy for reducing overfitting to a single simulated setting.
Why it matters: Simulation often makes reinforcement-learning experiments possible, but it is not a complete copy of reality. This concept prevents reading success in simulation as automatic real-world success.
Related concepts: simulation, real world, safety, deployment
Core Section: P4-19.3
Appears in: P4-19.4

AI service software regression¶

Meaning: AI service software regression is the quality drop where a function or behavior that used to work becomes worse after code, prompts, models, or settings change. In machine learning, regression can mean predicting a continuous value; software regression means a change has made existing behavior move backward in quality.
Why it matters: In an AI service, replacing a model or changing a prompt can make cases that used to work fail. This concept explains why repeatable checks and evaluation records are needed, and why before-and-after behavior must be compared under the same standard.
Related concepts: regression, metric, model validation, reproducibility
Core Section: P1-14.5
Appears in:

source data¶

Meaning: Source data is the starting record before it is redesigned into an analysis or learning problem.
Why it matters: Having source data is not the same as having a dataset or a learnable task. The same records can be reread into different sample units, feature tables, baselines, and output structures depending on the question.
Related concepts: dataset, data modeling, sample unit, feature
Core Section: P3-1.1
Appears in: P3-index, P3-1.2, P3-2.2, P3-3.1, P3-3.2, P3-4.1, P3-4.2, P3-4.3, P3-5.4, P3-5.6, P3-summary, P4-1.2

supervised learning label¶

Meaning: A label is the target answer or target output attached to a data example for the current task. It defines what the model is supposed to predict, such as a class name or numeric target.
Why it matters: Labels define supervised-learning targets, so changing the label definition changes the task and the evaluation. Poor or ambiguous labels can limit model quality before training even begins.
Related concepts: labeled example, data, model output, supervised learning, target, task definition
Core Section: P1-4.2
Appears in: P1-2.1, P4-index, P1-3.2, P1-8.1, P2-12.1, P2-12.2, P2-15.2, P3-3.1, P3-4.1, P3-4.2, P3-5.7, P3-8.6, P3-9.9, P3-9.10, P4-1.2, P4-2.1, P4-2.2, P4-2.3, P4-6.2, P4-6.4, P4-7.1, P4-12.1

SVM, support vector machine¶

Meaning: An SVM is a classification model that looks for a decision boundary between classes while making the margin to the closest cases as large as possible. For nonlinear problems, it can be extended by using a kernel to change the feature space in which the boundary is read.
Why it matters: SVM gives readers a representative model for reading boundary, margin, and kernel together. Beginners should not treat it only as an algorithm name to memorize, but as a way to ask what makes a classification boundary stable.
Related concepts: decision boundary, margin, kernel, feature space
Core Section: P4-13.1
Appears in: P4-13.2

supporting evidence¶

Meaning: Supporting evidence is information that supports, weakens, or refines a current claim or judgment.
Why it matters: Supporting evidence separates fluent output from supported output and makes review possible.
Related concepts: uncertainty, probability, probabilistic reasoning, provenance, human oversight
Core Section: P1-12.3
Appears in: P1-6.2, P1-10.3, P7-5.1, P7-5.3

sample unit¶

Meaning: A sample unit is the unit treated as one case for the current question, such as one row, one document, one event, or one model input.
Why it matters: The sample boundary determines dataset size, label meaning, feature interpretation, and how results are compared.
Related concepts: statistical sample, dataset, labeled example
Core Section: P3-4.1
Appears in: P2-11.2, P2-11.3, P2-12.3, P2-15.2, P3-1.1, P3-2.2, P3-3.2, P3-3.3, P3-4.2, P3-4.3, P3-5.1, P3-5.3, P3-5.4, P3-5.5, P3-5.6, P3-5.7, P7-index, P7-1.1, P7-1.3, P7-2.1, P7-3.2, P7-summary

sample space¶

Meaning: A sample space is the full set of possible outcomes for a probability problem.
Why it matters: Events and probabilities need a defined outer frame. Without a sample space, it is unclear what the possible outcomes are and what share of that whole an event represents.
Related concepts: outcome, event, probability
Core Section: P2-5.1
Appears in: P2-5.2

sampling bias¶

Meaning: Sampling bias is a state where the observed sample does not represent the population well and is skewed in a particular direction.
Why it matters: If the data used for learning or evaluation are biased, a model score can look better or worse than it would in the real target situation. The term helps readers ask whether the evaluation data are a reasonable proxy for future cases.
Related concepts: sample, bias, evaluation data, generalization
Core Section: P2-5.3
Appears in: P4-4.1

sampling¶

Meaning: Sampling is the procedure that selects one actual output piece from the candidate distribution computed by a model. After the model estimates which candidates are how plausible, sampling decides which candidate is actually taken. In a language model, this means choosing one of the next-token candidates; in image generation, it can mean following a probabilistic denoising path at the next step.
Why it matters: Sampling explains why the same input can produce different generated outputs, and why probability calculation and actual output selection must be separated. It also shows why generation quality is not determined only by model weights; the final choice procedure and settings such as temperature or top-k can strongly affect the result. Understanding sampling helps separate what the model has made plausible from what was actually selected.
Related concepts: next-token prediction, token, diffusion model
Core Section: P5-15.3
Appears in: P1-5.2, P1-10.2, P5-15.1, P5-15.2, P6-1.3, P6-4.1, P6-7.1, P6-7.2

scalar¶

Meaning: A scalar is a single numerical value. If a vector holds several ordered values, a scalar is one indivisible numeric result, such as a loss value, mean, or probability.
Why it matters: Learning and evaluation often reduce large tensor computations to single comparison values. Scalars help readers distinguish a full output from the one-number summary used for comparison or decision making.
Related concepts: vector, matrix, model output
Core Section: P2-3.1
Appears in: P2-3.2, P2-11.1, P2-15.1

sigmoid¶

Meaning: The sigmoid is an S-shaped function that compresses a real-valued score into a value between 0 and 1. It is often used when a score needs to be read as leaning toward one side of a binary decision.
Why it matters: It shows that a model score and a probability-like output are not the same layer of meaning. The score is transformed first, and any threshold or operational policy is applied afterward.
Related concepts: activation function, threshold, binary classification, logistic regression
Core Section: P2-2.4
Appears in: P5-3.2, P5-3.5, P5-3.6

candidate model score¶

Meaning: A candidate model score is a numeric value that a model assigns to a candidate output or decision target. A score can be used for comparison, ranking, or thresholding, but it does not always mean a calibrated probability.
Why it matters: Many inference results are not a single label; they include scores for candidates or cases. This concept helps separate model output, probability interpretation, and the final business decision, so a high score is not mistaken for an already finalized action.
Related concepts: probability, prediction, modeling task
Core Section: P1-5.2
Appears in: P2-11.1, P2-11.2, P2-11.3, P2-11.4, P3-9.8, P3-9.12, P4-6.2, P4-6.4

state-space search¶

Meaning: State-space search is the process of following possible states, actions, or candidate paths to find a goal or solution. It is not applying one formula at once; it is choosing which candidate path to inspect next.
Why it matters: Search is a core AI problem-solving pattern behind path finding, game moves, planning, combinatorial optimization, and candidate retrieval. It connects directly to search space, computational limits, heuristics, and pruning.
Related concepts: heuristic, knowledge representation, search space, computational limit
Core Section: P1-7.1
Appears in: P1-13.2, P1-13.3, P1-13.4, P1-16.2

search index¶

Meaning: A search index is a structure prepared in advance so relevant vectors or documents can be found faster.
Why it matters: Large-scale retrieval cannot compare every candidate from scratch each time. Index design affects speed, memory use, recall, and the practical quality of vector search services.
Related concepts: similarity search, ANN, approximate nearest neighbor, vector database
Core Section: P1-13.4
Appears in: P6-13.2

search space¶

Meaning: A search space is the full structure of states, choices, and candidate paths that may be considered while solving a problem. It includes where the system is, what can be done next, and what new states those choices can lead to.
Why it matters: A search space explains why candidate counts grow quickly as choices accumulate and why checking every candidate becomes difficult. It helps readers see that slow computation may come from the size of the candidate structure itself, not just from poor implementation.
Related concepts: search, computational limit, heuristic
Core Section: P1-7.1
Appears in: P1-7.2, P1-7.4, P2-summary

security¶

Meaning: Security is the condition that protects an AI system from misuse or damage across inputs, outputs, permissions, connected tools, and stored data paths.
Why it matters: Generative AI systems often connect to retrieval, code execution, file changes, and external APIs. Security helps readers separate answer quality failures from dangerous execution, permission abuse, and sensitive information exposure.
Related concepts: privacy, sensitive information, permission, least privilege, accountability
Core Section: P1-15.3
Appears in: P1-14.2, P1-14.4, P1-14.6, P1-16.1

selective labels¶

Meaning: Selective labels are labels left only on some cases, usually because those cases were reviewed by a person or passed through a prior decision rule. The observed labels may therefore reflect the review path rather than a random sample of all events.
Why it matters: If a model or cause distribution is evaluated only on selectively labeled cases, errors in the unreviewed range can remain hidden. This concept separates has a label from represents the whole, and requires the meaning of missing labels, the review path, and possible bias to be written together.
Related concepts: supervised learning label, review queue, bias, provenance
Core Section: P3-8.6
Appears in: P3-8.6, P3-8.7

self-attention¶

Meaning: Self-attention is an attention mechanism in which each token in the same sequence calculates relationships with other tokens and updates its own representation. Instead of looking only at the current position, it mixes information from other positions with learned weights.
Why it matters: Self-attention is a core mechanism that lets Transformers compute contextual relationships directly without sequential recurrence. It helps explain why a token representation can reflect other tokens in the same sequence and why Transformers can mix relationships among many positions within one layer.
Related concepts: Transformer, Attention, token
Core Section: P5-13.2
Appears in: P1-11.3, P5-14.1, P5-14.2, P6-4.1

sensitive information¶

Meaning: Sensitive information is information that can directly harm a person, organization, transaction, or security posture if exposed. It includes personal data, secrets, internal documents, credentials, and other high-risk material.
Why it matters: AI systems can copy the same information through prompts, retrieval documents, logs, evaluation samples, and debugging records. Sensitive information requires input limits, least privilege, masking, and review controls.
Related concepts: privacy, security, provenance, permission, human oversight
Core Section: P1-15.3
Appears in: P1-10.3, P1-14.2, P1-16.2

similarity¶

Meaning: Similarity is a comparison criterion for how close two samples, vectors, or representations are. It may be based on distance, direction, shared features, or another task-specific rule.
Why it matters: Clustering and retrieval both depend on what is counted as similar. Understanding similarity keeps the reader from treating a high similarity score or a cluster assignment as an automatic fact; it is a result under a chosen comparison rule.
Related concepts: distance, cosine similarity, similarity search, vector space
Core Section: P1-13.2
Appears in: P1-13.2, P2-3.2, P4-17.1, P6-3.2

similarity search¶

Meaning: Similarity search finds vectors near the vector representation of a question or document and selects related candidates. Instead of looking only for exact string matches, it searches a representation space for candidates that are semantically close.
Why it matters: Similarity search is a core step in RAG because it connects embeddings to an actual retrieval flow. It also shows that keyword search and vector search are not simply rivals; some problems need both to retrieve stable evidence. Its strength is finding nearby meaning candidates rather than only exact wording.
Related concepts: embedding, retrieval-augmented generation, RAG, vector, similarity, top-k
Core Section: P1-13.2
Appears in: P1-13.3, P1-13.4, P6-3.1

softmax¶

Meaning: Softmax turns several scores into normalized values that sum to 1, so they can be read like a probability distribution across candidates or classes.
Why it matters: Softmax connects class scores, probability-like outputs, and final candidate selection. It explains why all candidate scores must be compared together and why a class probability is relative to the other class scores.
Related concepts: exponential function, logarithm, classification, logistic regression
Core Section: P2-2.4
Appears in: P4-11.4, P5-3.6, P6-4.1

sparsity¶

Meaning: Sparsity is the problem where observed cases are too thin compared with the possible combinations of words, states, or features, making probabilities or weights hard to estimate reliably.
Why it matters: Sparsity explains why early count-based language models hit limits as context grows and why generalization and vector representations become important. Many possible cases remain unseen even when the dataset looks large.
Related concepts: corpus, statistical language model, language model
Core Section: P1-11.1

speech generation¶

Meaning: Speech generation is a task that creates audio signals over time. It is not just choosing text tokens; it builds audible waveforms or acoustic features that unfold along a time axis.
Why it matters: Speech generation shows that generation is not limited to text or images. It helps readers see how deep learning extends to continuous sequential outputs, including but not limited to text-to-speech.
Related concepts: generation, sequence modeling
Core Section: P1-9.2

standard deviation¶

Meaning: Standard deviation is the square root of variance. It shows how far values typically spread around the mean in units closer to the original data.
Why it matters: Variance can be hard to interpret because it uses squared units. Standard deviation makes spread easier to compare, so readers can distinguish a large value from a highly variable set of values.
Related concepts: variance, distribution, standard error
Core Section: P2-5.5
Appears in: P3-1.2, P4-12.2

standardization¶

Meaning: Standardization transforms each feature by subtracting its mean and dividing by its standard deviation so different units and spreads become more comparable. In distance-based models, it can be read as preprocessing that rebalances how much each feature influences distance.
Why it matters: Without standardization, a large-range feature can dominate distance calculation. After standardization, smaller-range features can affect neighbor selection again, so the before-and-after neighbor composition needs to be inspected.
Related concepts: feature, distance, preprocessing
Core Section: P4-12.2
Appears in: P4-12.2

agent state¶

Meaning: Agent state is a summary of the current situation that the next action can use. In reinforcement learning it describes the environment's current situation, and in AI-agent execution it can include gathered materials, tool results, and intermediate judgments.
Why it matters: Search, reinforcement learning, and AI-agent execution all choose next actions based on what has already been reached or observed. State is not just a record; it is the representation that the next decision can use.
Related concepts: action, reinforcement learning environment, reinforcement learning policy, observation
Core Section: P1-7.1
Appears in: P1-8.3, P1-14.3, P1-14.4, P1-14.5, P5-12.1, P5-12.2, P7-6.1, P7-6.2, P4-2.3

statistical language model¶

Meaning: A statistical language model estimates the likelihood of the next word, token, or expression from frequencies and conditional probabilities observed in a corpus.
Why it matters: Statistical language models show that language was treated as a computable probability problem before modern LLMs. They provide a bridge from frequency-based prediction to distributed representations and neural language models.
Related concepts: language model, sparsity, distributed representation
Core Section: P1-11.1
Appears in: P1-9.3, P1-13.2, P1-13.3, P6-20.1

stochastic process¶

Meaning: A stochastic process is a process whose evolution includes probabilistic variation over time. It is not just one random event, but a sequence of states or outcomes that can unfold probabilistically.
Why it matters: It separates using probability as a number from modeling a process that changes probabilistically over time. The concept helps interpret reinforcement-learning environments, sensor noise, and time-series variation.
Related concepts: probability, random, nondeterminism
Core Section: P1-6.2

semi-supervised learning¶

Meaning: Semi-supervised learning is a learning setup that uses a small amount of labeled data together with a larger amount of unlabeled data.
Why it matters: It becomes relevant when labels are expensive or scarce. Unlabeled structure can help guide learning, but clusters or neighbors should not be spread as correct labels without review because mistakes can spread with them.
Related concepts: unsupervised learning, clustering, supervised learning label
Core Section: P4-17.2
Appears In: P4-17.4

supervised learning¶

Meaning: Supervised learning is a learning setup that uses input examples together with human-provided answer labels to learn a relationship from inputs to outputs. The target of prediction is defined before training.
Why it matters: It gives beginners the basic frame that connects examples, features, labels, learning, and generalization. Understanding supervised learning also makes the contrast with clustering clear: supervised learning predicts known targets, while clustering explores structure without answer labels.
Related concepts: supervised learning label, model training, generalization
Core Section: P4-2.1
Appears in: P1-3.2, P4-index, P1-8.1, P4-8.1, P4-10.1, P4-11.1, P4-12.1, P4-13.1, P4-14.1, P4-15.1, P4-16.1, P4-17.1, P4-2.1, P4-2.2, P4-2.3

symbolic AI¶

Meaning: Symbolic AI is an AI approach that uses symbols, rules, logic, and explicit knowledge representation to solve problems. Instead of learning statistical patterns from many examples, it writes concepts and rules in forms that can be manipulated.
Why it matters: Symbolic AI is a major historical AI tradition before modern machine learning and remains important for understanding rule-based systems, expert systems, explicit knowledge, and explainable rule application. It keeps learning from data separate from applying written rules.
Related concepts: rule-based system, knowledge representation
Core Section: P1-2.1
Appears in: P1-index, P1-2.3, P1-3.1, P1-8.1, P1-11.1, P1-summary

sequence modeling¶

Meaning: Modeling data whose order matters, such as text, time series, audio, or action traces.
Why it matters: It explains why some models must preserve temporal or positional relationships instead of treating examples as unordered.
Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
Core Section: P1-9.3
Appears in: P1-11.1, P1-11.2, P1-11.3

statistical inference¶

Meaning: The process of drawing conclusions about a population or process from sample data while accounting for uncertainty.
Why it matters: It distinguishes evidence-based estimation from simply describing the sample at hand.
Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
Core Section: P1-5.3

statistical sample¶

Meaning: A subset of a population or process observed for analysis or estimation.
Why it matters: It matters because conclusions from data depend on whether the sample represents the larger situation being discussed.
Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
Core Section: P2-5.3
Appears in: P1-17.2, P2-5.4, P2-6.1