Concept Glossary: I¶

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

image recognition¶

• Meaning: Image recognition is the task of predicting what is present in an input image or which category the image belongs to. It connects arrays of pixels to meaningful visual categories.
• Why it matters: Image recognition is a clear example of the shift from hand-crafted visual cues to learned representations. It helps readers see that image input may be divided into classification, detection, segmentation, and other more specific visual tasks.
• Related concepts: CNN, convolutional neural network, feature, learned representation
• Core Section: P1-9.1
• Appears in: P1-9.2, P1-9.3, P1-10.1, P5-2.1, P5-11.1, P5-11.2

in-context learning¶

• Meaning: In-context learning is the behavior or usage pattern in which a model changes its output behavior based on instructions and examples in the prompt, without updating model weights.
• Why it matters: This concept explains why natural-language instructions and in-prompt examples matter in LLM use. It also separates changing the prompt from fine-tuning or retraining the model.
• Related concepts: pretraining, fine-tuning, GPT
• Core Section: P1-12.2
• Appears in: P1-11.3

input context¶

• Meaning: Input context is the background information, document material, and situation provided inside the model input so a task or response can be interpreted correctly. It changes how the same instruction should be read.
• Why it matters: The same request can produce different useful answers depending on the audience, source material, and scope included as input context. Context selection is therefore a design choice in prompting, RAG, and long-document work.
• Related concepts: prompt, prompt structuring, in-context learning
• Core Section: P1-12.1

incomplete information¶

• Meaning: Incomplete information is a situation where the facts needed for judgment or choice are not fully available at the start. The rules or goal may exist, but some necessary input facts are missing.
• Why it matters: When key information is missing, a system may need probability, search, retrieval, or human review rather than just more rules. The concept helps readers distinguish lack of information from lack of logic.
• Related concepts: partial observability, uncertainty, rule-based system
• Core Section: P1-6.1
• Appears in: P1-2.2, P1-7.1

inference¶

• Meaning: Inference has multiple meanings. In symbolic AI, it can mean deriving a conclusion from known facts and rules. In machine learning, model inference often means running a trained model on new input to obtain output.
• Why it matters: The same word can refer to reasoning, prediction, or model execution depending on context. Readers need this distinction to avoid mixing a logical conclusion process with the runtime use of a trained model.
• Related concepts: inference engine, model training, prediction, generation
• Core Section: P1-5.3
• Appears in: P1-2.1, P1-3.2, P1-5.2, P5-6.1, P5-6.3

inference engine¶

• Meaning: An inference engine is the component that finds and applies rules that match the current facts. It turns stored rules into an execution process that can derive conclusions.
• Why it matters: Rule-based AI is not just a list of rules. The inference engine shows how knowledge representation and rule-application procedure work together to produce a result.
• Related concepts: rule-based approach, knowledge base
• Core Section: P1-3.1
• Appears in: P1-2.1

information retrieval¶

• Meaning: Information retrieval is the task of finding documents, passages, or evidence candidates that match a current question or information need. It is often a ranking problem rather than a simple exact string match.
• Why it matters: Retrieval-augmented generation depends on retrieving relevant evidence before generation. Weak retrieval can make the generated answer weak even when the language model is fluent.
• Related concepts: retrieval-augmented generation, RAG, similarity search, search, ranking
• Core Section: P1-13.3
• Appears in: P1-13.4, P1-14.2, P7-5.1, P7-5.3

instruction tuning¶

• Meaning: Instruction tuning is an additional adaptation stage that helps a pretrained model respond better to natural-language user instructions and expected answer formats.
• Why it matters: A model can know language patterns without automatically following requests well. Instruction tuning helps explain why chat-style LLMs behave more like assistants than plain continuation models.
• Related concepts: fine-tuning, pretraining, prompt
• Core Section: P6-10.1
• Appears in: P6-10.2

intermediate representation¶

• Meaning: An intermediate representation is a human-designed representation placed between raw data and final features so structure, order, or patterns become easier to inspect.
• Why it matters: Some problems are hard to see from raw logs or simple numeric summaries. Intermediate representations let readers compare whether segments, token sequences, or pattern blocks reveal the problem better.
• Related concepts: feature, representation, representation learning
• Core Section: P3-6.2
• Appears in: P3-index, P3-5.1, P3-5.3, P3-6.1, P3-6.3, P3-summary

interpretation boundary¶

• Meaning: An interpretation boundary marks how far a data result or comparison can be stated, and where the explanation must stop until more evidence is available.
• Why it matters: A comparison table can show a change signal without proving a cause, diagnosis, or future outcome. Interpretation boundaries keep conservative outputs such as needs review separate from stronger claims such as confirmed anomaly.
• Related concepts: interpretation, comparison report, output structure, review queue
• Core Section: P3-8.2
• Appears in: P3-1.2, P3-8.4, P3-8.7, P3-9.1, P3-summary

intervention feedback¶

• Meaning: A feedback structure where model outputs, review rules, or operational actions change later data and labels, so the results observed later differ from the original natural course.
• Why it matters: If intervention reduces later events or changes log length and label selection, later data may not prove that the cases were originally safe. It may instead be the result left after action came first. This concept helps separate pre-intervention signals, post-intervention operational results, and selectively observed labels.
• Related concepts: interpretation boundary, selective labels, review queue, provenance
• Core Section: P3-8.7
• Appears in: P3-8.7

information integrity¶

• Meaning: The reliability and trustworthiness of information as it is created, transmitted, transformed, and reused.
• Why it matters: It matters because AI outputs, retrieved sources, and summaries can degrade if provenance, context, or factual consistency is lost.
• Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
• Core Section: P1-10.3
• Appears in: P1-15.1