Concept Glossary: O¶

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

object detection¶

• Meaning: Object detection is a vision task that predicts not only what is in an image, but also where each object is. It does not stop at there is a cat; it also asks where the cat is and how much space it occupies. The output is usually a set of category + location results for each detected object, not a single label.
• Why it matters: Object detection has a more complex output structure than image classification, and it is a representative example of deep learning expanding from category prediction to category and location prediction together. It helps separate classification, detection, and segmentation even when they all use image inputs. Understanding object detection also makes it easier to see why multi-step vision pipelines can be reframed as one prediction problem.
• Related concepts: bounding box, end-to-end learning, image recognition, output structure
• Core Section: P1-9.2
• Appears in: P1-9.3, P1-10.1, P5-11.1, P6-20.2

objective function¶

• Meaning: An objective function is the overall criterion that a learning or optimization process tries to decrease or increase. It can combine sample-level loss, average loss over many samples, regularization penalties, and constraints into one rule for deciding what counts as better. A loss function can be a central component inside an objective function, but the objective function is the broader scoring rule.
• Why it matters: A model does not move toward being vaguely better; it is adjusted in the direction defined by the objective function. This concept connects loss functions, regularization, and constraints into the actual learning target. It also separates evaluation statements such as the accuracy is high from the quantity that training directly optimized.
• Related concepts: loss function, optimization, metric
• Core Section: P2-6.2
• Appears in: P2-6.3, P2-15.2

OOB score(out-of-bag score)¶

• Meaning: An internal validation-like score in random forest that uses samples not drawn into a tree's bootstrap sample.
• Why it matters: Because each tree leaves some samples out of its bootstrap draw, random forest can get a rough self-check signal during training. This score should be read as a quick inspection aid, not as a replacement for final testing.
• Related concepts: random forest, bootstrap, validation data, metric
• Core Section: P4-15.3
• Appears in: P4-15.4

observation¶

• Meaning: An observation is new information or a result obtained after taking an action. In reinforcement learning it may be the next screen, number, or signal returned by the environment. In a service-style AI agent, it can be a search result, error message, file content, or test output used for the next decision.
• Why it matters: In multi-step problems, what is learned after an action can immediately change the next action or the decision to stop. Search results, error messages, tool outputs, and test results all update state and reshape the plan. This concept helps frame AI agent execution as a repeated cycle of action, observation, and selection rather than a single answer.
• Related concepts: AI agent, state, action
• Core Section: P1-14.3
• Appears in: P1-14.4, P1-14.5

optimization¶

• Meaning: Optimization is the problem of finding a better value among many candidates while considering criteria and constraints. It does not always mean computing one perfect mathematical answer. It often refers to the whole process of adjusting values in a direction that improves the current objective function.
• Why it matters: Optimization lets readers understand learning as a search for values that reduce loss, not as memorizing answers. It gives a shared frame for model training, path search, and resource allocation: what counts as good, under what constraints, and which values are being adjusted. It also explains why learning is usually expressed as an iterative process.
• Related concepts: objective function, gradient descent, constraint
• Core Section: P2-6.1
• Appears in: P2-6.2, P2-6.3, P2-15.2, P4-3.1, P4-7.2, P4-7.3

optimizer¶

• Meaning: A learning algorithm that decides how model parameters are updated to reduce loss.
• Why it matters: Even with the same loss function and data, different update rules can change learning speed, stability, and convergence behavior. Understanding optimizers makes learning rate, gradients, and Adam readable as mechanisms that shape the learning path, not just as settings.
• Related concepts: optimization, gradient descent, learning rate, backpropagation, loss function
• Core Section: P5-7.1
• Appears in: P5-6.1, P5-7.2, P5-7.3

outlier¶

• Meaning: An outlier is a value that appears unusually far from the overall pattern of values. Being visibly different does not automatically mean the value is wrong. It may be an input mistake, a rare real event, or an important warning signal.
• Why it matters: Outliers can come from input errors, measurement problems, or rare but meaningful cases, so they must be interpreted before being discarded. This concept makes readers consider how extreme values affect interpretation and model training, not only the average pattern near the center of a distribution. In small datasets, only a few outliers can strongly shift means, variances, and regression lines.
• Related concepts: distribution, visualization
• Core Section: P2-13.1
• Appears in: P2-13.2, P3-5.2, P4-2.2, P4-10.2, P4-10.3

modeling output structure¶

• Meaning: Modeling output structure is the designed result frame that determines how a computation should be delivered, such as a comparison report, review candidate queue, or target label candidates. It asks not only what is the answer?, but also what container should hold the result, and who will use it next? Choosing between a single number, a human-readable comparison table, or a candidate set for later training is a choice of output structure.
• Why it matters: The same source data leads to different dataset designs depending on whether the result becomes a comparison table, a review workflow, or future training candidates. Output structure must be decided before review procedures and automation steps can be designed. This concept separates what answer should be produced? from what workflow-ready form should carry that answer?
• Related concepts: comparison report, target, sample unit, task definition
• Core Section: P3-2.2
• Appears in: P3-index, P3-1.1, P3-1.2, P3-1.3, P3-2.1, P3-3.2, P3-4.3, P3-8.2, P3-8.4, P3-9.1, P3-9.2, P3-9.3, P3-9.5, P3-summary, P7-index, P7-1.1, P7-2.1, P7-5.1, P7-summary

overfitting¶

• Meaning: Overfitting is the condition in which a model fits the training data too closely and performs worse on new data. The model has learned not only real repeated patterns, but also accidental fluctuations or noise that appeared only in the training set.
• Why it matters: Overfitting is central to understanding why learning from data is not the same as memorization. It explains why training scores and validation scores are checked separately, and why regularization and validation-based stopping become important as models grow more complex. It also separates high training performance from useful performance on new cases.
• Related concepts: generalization, underfitting, regularization
• Core Section: P4-5.1
• Appears in: P1-3.2, P4-3.2, P4-4.1, P4-5.2, P4-6.1, P4-9.1, P4-9.2, P4-14.2, P4-16.2, P5-8.1, P5-8.2

one-hot representation¶

• Meaning: A sparse representation in which one position is 1 and the remaining positions are 0 to identify a category.
• Why it matters: It provides a simple contrast to dense embeddings, which can represent similarity rather than only identity.
• Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
• Core Section: P1-11.1

open-weight model¶

• Meaning: A model whose trained weight files are made available so users can download, run, or adapt the model in their own environment. Open weights do not automatically mean that training data, training code, evaluation procedures, or usage rights are fully open.
• Why it matters: Open-weight models give users more control over hosting, cost structure, data location, and adaptation. They also move license review, hardware planning, safety filtering, updates, and evaluation responsibility closer to the user.
• Related concepts: weight, parameter, license, inference
• Core Section: P6-21.1
• Appears in: