Concept Glossary: E¶

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

embedding¶

• Meaning: An embedding turns text, tokens, sentences, or documents into vectors that models and search systems can compare.
• Why it matters: Embeddings are the basis for semantic comparison, vector search, and RAG, so they connect language data to computable similarity.
• Related concepts: distributed representation, similarity search, vector, representation
• Core Section: P6-3.1
• Appears in: P1-11.1, P1-13.1, P1-13.2, P1-13.3, P5-10.1, P6-3.2, P6-3.3, P6-3.4, P7-5.1

encoder¶

• Meaning: An encoder reads an input and turns it into internal representations that reflect context.
• Why it matters: It separates the role of reading and representing input from the role of generating output. This distinction helps compare encoder-centered models, decoder-centered models, and encoder-decoder structures.
• Related concepts: decoder, Encoder-Decoder, contextual representation
• Core Section: P1-11.3
• Appears in: P1-11.2, P6-6.1, P6-6.2

Encoder-Decoder¶

• Meaning: An Encoder-Decoder separates reading an input sequence from generating an output sequence.
• Why it matters: It explains why translation and summarization models can be read as two linked roles: first encoding the source sequence, then decoding a new sequence from that representation. It also makes the fixed-length bottleneck and the later need for attention easier to understand.
• Related concepts: encoder, decoder, Attention
• Core Section: P1-11.2
• Appears in: P1-11.3, P6-5.1

ensemble¶

• Meaning: An ensemble combines predictions from multiple models into one final judgment. It may use averaging, voting, or sequential correction to obtain a result that is more stable or stronger than one model alone.
• Why it matters: Ensemble methods may average independently varied models, collect votes, or add sequential corrections to earlier errors. The key question is not only how many models are used, but how their predictions are combined.
• Related concepts: decision tree, overfitting, error
• Core Section: P4-15.1
• Appears in: P4-15.1, P4-16.1

reinforcement learning environment¶

• Meaning: A reinforcement learning environment is the outside world or interaction target that responds to a reinforcement learning agent's actions with state changes and rewards.
• Why it matters: This concept gives readers the vocabulary to separate roles, boundaries, and failure modes that otherwise get mixed together in AI systems and data work.
• Related concepts: AI agent, state, reward
• Core Section: P1-8.3
• Appears in: P4-2.3

error¶

• Meaning: An error is the difference between an estimate or prediction and the actual value.
• Why it matters: This concept gives readers the vocabulary to separate roles, boundaries, and failure modes that otherwise get mixed together in AI systems and data work.
• Related concepts: estimation, loss, sample
• Core Section: P2-5.3
• Appears in: P2-5.4, P2-6.1, P2-6.3, P2-15.1, P4-10.2

error cost¶

• Meaning: Error cost is the real operational loss or burden created when a model is wrong. Different errors can have different costs, such as missing a risky case, adding unnecessary review work, or triggering the wrong automatic action.
• Why it matters: Even with the same accuracy, the better threshold and review-queue policy can change depending on which error is more costly. This concept turns false negatives and false positives from simple counts into a question about what the current problem is trying harder to reduce.
• Related concepts: false negative, false positive, threshold, review queue
• Core Section: P3-9.12
• Appears in: P3-9.12

estimation¶

• Meaning: Estimation uses observed samples to infer a value or property of a broader population.
• Why it matters: This concept gives readers the vocabulary to separate roles, boundaries, and failure modes that otherwise get mixed together in AI systems and data work.
• Related concepts: population, sample, error
• Core Section: P2-5.3
• Appears in: P2-5.4, P2-6.1

evaluation data¶

• Meaning: Evaluation data is data kept out of direct model training and used to check how the trained model behaves. In this entry, the term is used broadly before separating validation data and test data more finely.
• Why it matters: To check whether a model only fits data it has already seen, or whether it can generalize to unseen cases, the check must use data not used for training. The term helps readers avoid mixing the training score with an estimate of new-data performance.
• Related concepts: training data, validation data, test data, generalization, metric
• Core Section: P4-4.1
• Appears in: P4-4.1, P4-4.2

model evaluation design¶

• Meaning: Model evaluation design is the choice of data split, metric, and comparison condition used to evaluate a model or rule. It is not only calculating a score, but first deciding the test conditions that fit the problem structure.
• Why it matters: The same model can be judged very differently depending on whether the problem needs a time split, group split, ranking metric, or regression metric. If evaluation design does not match the problem structure, a high score may fail to describe real operational performance.
• Related concepts: time split, group split, data leakage, ranking
• Core Section: P3-9.13
• Appears in: P3-9.13

probability event¶

• Meaning: A probability event is a set of outcomes grouped by a condition of interest inside a sample space.
• Why it matters: This concept gives readers the vocabulary to separate roles, boundaries, and failure modes that otherwise get mixed together in AI systems and data work.
• Related concepts: outcome, sample space, probability
• Core Section: P2-5.1
• Appears in: P2-5.2, P2-5.3

excessive agency¶

• Meaning: Excessive agency is a state where an AI agent or automation system has broader action authority than the task requires.
• Why it matters: This concept gives readers the vocabulary to separate roles, boundaries, and failure modes that otherwise get mixed together in AI systems and data work.
• Related concepts: least privilege, permission, human oversight, security, AI agent
• Core Section: P1-15.3
• Appears in: P1-10.3, P1-14.6, P7-6.2

expert system¶

• Meaning: An expert system is a rule-based AI system that encodes domain expert judgment as rules and knowledge.
• Why it matters: Expert systems are a clear historical example of what rule-based AI could do well and where it became hard to maintain. They show both the strength of explicit knowledge and the cost of encoding expert judgment by hand.
• Related concepts: rule-based system, knowledge base, knowledge representation, inference engine
• Core Section: P1-3.1
• Appears in: P1-2.1, P1-2.2

exploitation¶

• Meaning: Exploitation chooses the action that currently looks best from information already gathered.
• Why it matters: It forms one side of the reinforcement-learning tradeoff with exploration. Too little exploitation wastes what has already been learned, while too much can prevent the learner from discovering better long-term actions.
• Related concepts: exploration, reinforcement learning policy, reward
• Core Section: P1-8.3
• Appears in: P4-2.3

exploration¶

• Meaning: Exploration deliberately tries less-tested actions to gather more information.
• Why it matters: Without exploration, a learner can miss better long-term actions; without exploitation, it may fail to use what it has learned.
• Related concepts: exploitation, reinforcement learning policy, reward
• Core Section: P1-8.3
• Appears in: P4-2.3, P4-19.3, P4-summary

error accumulation¶

• Meaning: The process by which small errors build up across repeated steps, stages, or generated outputs.
• Why it matters: It matters because sequence generation, long workflows, and multi-step analysis can fail gradually even when each individual step looks acceptable.
• Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
• Core Section: P1-10.2
• Appears in: P1-10.3

end-to-end learning¶

• Meaning: A learning setup in which a model is trained to connect raw or lightly processed inputs directly to the final target output.
• Why it matters: It explains why some deep learning systems replace hand-designed intermediate stages with a single jointly trained pipeline.
• Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
• Core Section: P1-9.2

exponential function¶

• Meaning: A function in which the variable appears in the exponent, producing multiplicative growth or decay.
• Why it matters: It appears repeatedly in probability, optimization, neural networks, and log/exp transformations.
• Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
• Core Section: P2-2.4