Concept Glossary: B¶

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

backpropagation¶

• Meaning: Backpropagation is the procedure that computes how much the loss at the output depends on each parameter by working backward through the network. More precisely, it applies the chain rule along the computation graph to distribute the final error signal across intermediate activations and weights.
• Why it matters: A neural network cannot reduce loss unless it knows which direction each weight should move. Backpropagation provides that gradient signal efficiently, so learning becomes a concrete calculation rather than a vague idea that the network adjusts itself. It also explains why deep layers can be trained by tracing each connection's contribution to the final error.
• Related concepts: gradient, loss function, chain rule
• Core Section: P5-5.1
• Appears in: P5-5.2, P5-6.1, P5-7.1

baseline¶

• Meaning: A baseline is a reference point used to read change or difference by comparing the current value with a normal range or another target. It keeps a number from being interpreted by itself and asks what it changed against.
• Why it matters: A recent value or model score is hard to interpret alone. A baseline gives comparison reports and operational data a stable point of reference, so readers ask how much something improved, worsened, or moved from the expected pattern instead of overreading an isolated number.
• Related concepts: comparison report, output structure, baseline model, evaluation
• Core Section: P3-7.3
• Appears in: P3-1.1, P3-1.2, P3-2.2, P3-3.2, P3-4.1, P3-5.1, P3-7.1, P3-7.4, P3-7.5, P3-8.1, P3-9.1, P3-9.3, P3-summary, P4-8.2, P7-index, P7-1.1, P7-1.3, P7-2.2, P7-7.3, P7-summary

baseline model¶

• Meaning: A baseline model is the simplest comparison model or score used to check whether a more complex model has made a meaningful improvement. It gives the project an interpretable starting point before adding more complexity.
• Why it matters: A high score can come from an easy task rather than a better modeling method. A baseline model fixes the minimum comparison point that candidate models must beat, and it makes error examples and execution summaries more useful because readers can see what changed beyond the simplest approach.
• Related concepts: evaluation, baseline, model selection, task definition
• Core Section: P4-8.2
• Appears in: P4-index, P4-6.1, P4-6.2, P4-9.1, P4-10.1, P4-12.1, P4-13.1, P4-15.1, P4-16.1, P4-summary, P7-index, P7-2.1, P7-4.2, P7-summary

baseline window¶

• Meaning: A baseline window is the time period or observation range treated as the normal state for comparison with recent change. It is not every old data point; it is the reference interval chosen for the current question.
• Why it matters: Even when a baseline exists, interpretation becomes unstable if the reference period is unclear. The baseline window makes readers check sample size, seasonality, event effects, and channel mix before saying that a current value improved, worsened, or moved outside the usual range.
• Related concepts: baseline, sample, comparison table, indicator, retrospective
• Core Section: P7-1.3
• Appears in: P7-index, P7-7.3, P7-7.4, P7-summary

batch¶

• Meaning: A batch is a group of items processed together. In operations it can mean grouping many requests or jobs; in deep learning it means a group of samples computed at once. The shared idea is treating multiple items as one processing unit.
• Why it matters: Batches affect throughput and cost in services, and they are a basic computation unit in deep learning. Batch size changes memory use, gradient noise, and training speed, so a batch is both an operational grouping and a learning-procedure design choice.
• Related concepts: throughput, operation, tensor
• Core Section: P1-14.6
• Appears in: P5-6.1, P5-9.2

batch normalization¶

• Meaning: Batch normalization normalizes activation distributions by using the mean and variance of a batch so the next layer receives values in a more manageable range. It is an internal stabilization step, not only input preprocessing.
• Why it matters: In deep networks, shifting activation distributions can make training unstable. Batch normalization helps stabilize computation and also explains why some layers behave differently in training mode and evaluation mode, where current batch statistics and accumulated statistics are handled differently.
• Related concepts: training mode, evaluation mode, numerical stability
• Core Section: P5-8.3
• Appears in: P5-6.3, P5-8.1

BERT¶

• Meaning: BERT is a representative language-model family pretrained with a Transformer encoder to build representations that consider the whole input context. It is better understood as a model for contextual representation than as a model that mainly continues text from left to right.
• Why it matters: BERT is a useful contrast with GPT because models can share the Transformer family while serving different purposes. Encoder-centered models are often used for understanding-oriented tasks such as sentiment classification, sentence-pair judgment, and representation extraction for search.
• Related concepts: encoder, pretraining, contextual representation
• Core Section: P6-20.1
• Appears in: P1-11.3, P6-20.2

bias¶

• Meaning: Bias is the problem where an AI system repeatedly produces more harmful, unfavorable, or distorted results for certain people, groups, or situations. It is not just a single wrong prediction; it asks who receives failures more often and in what direction.
• Why it matters: Bias can arise from data collection, labeling standards, feature choices, evaluation methods, and deployment context, even without bad intent. A model can look strong on average while still creating concentrated harm, so group-level error patterns and operating context must be examined separately.
• Related concepts: safety, accountability, human oversight
• Core Section: P1-15.1

bias term¶

• Meaning: A bias term is a model parameter that adjusts the default position of an output even when input signals are weak or zero. Like the intercept in a line, it moves the starting level of the output separately from the input weights.
• Why it matters: Weights explain how inputs affect the output, but they do not by themselves set the output's default starting point. The bias term helps readers see that training adjusts both input effects and baseline output position, which matters in linear models and neural-network layers.
• Related concepts: weight, parameter, training
• Core Section: P1-5.1

boolean mask¶

• Meaning: A boolean mask selects values by using a true-or-false array of the same length or compatible shape. Instead of naming positions directly, it first evaluates a condition and then keeps the elements where the condition is true.
• Why it matters: Boolean masks support condition-based selection in preprocessing and array work. They differ from simple slicing because the result shape and copying behavior can change, so readers need to understand them before interpreting filtered samples or conditional array operations.
• Related concepts: filtering, fancy indexing, copy
• Core Section: P2-11.4
• Appears in: P2-12.2

bounding box¶

• Meaning: A bounding box represents an object's position in an image as a rectangular region. It records where the object is, often with corner coordinates or a center point plus width and height.
• Why it matters: Object detection is not only about saying what is present; it must also say where it is. Bounding boxes make this output structure visible by combining location and class information, while remaining different from pixel-level segmentation that tries to describe the exact object boundary.
• Related concepts: object detection, YOLO, image recognition
• Core Section: P1-9.2
• Appears in: P1-10.1

BPE, Byte Pair Encoding¶

• Meaning: Byte Pair Encoding is a tokenizer method that builds a subword vocabulary by repeatedly merging character pieces or fragments that often appear together. It starts from small units and gradually treats frequent combinations as larger token units.
• Why it matters: BPE shows why tokenization is not simple whitespace splitting. The learned vocabulary affects token counts, cost, and how rare or frequent words are divided, so the length a user sees and the length a model computes can differ.
• Related concepts: tokenization, WordPiece, SentencePiece
• Core Section: P6-2.2
• Appears in: P6-2.2, P7-4.1

branch¶

• Meaning: A branch is a named line of history that lets work proceed separately inside the same repository. It is not just a copied folder; it separates change history and readiness for publication within one project.
• Why it matters: Branches let a project keep draft work, experimental changes, and deployable content apart. The concept helps readers understand that the same file can have a different role depending on which branch it belongs to, especially when one branch is tied to publication.
• Related concepts: commit, deployment, workflow
• Core Section: P2-14.2
• Appears in: P2-14.1, P7-7.1

broadcasting¶

• Meaning: Broadcasting is a NumPy computation rule that lets a compatible smaller array or scalar be applied across a larger array. It is better understood as a shape interpretation rule than as manually copying values into a huge array.
• Why it matters: Broadcasting explains how short array expressions can apply values along particular axes without explicit loops. It helps readers debug cases where an operation works syntactically but spreads values along a different axis than expected.
• Related concepts: vectorization, shape, scalar
• Core Section: P2-11.3
• Appears in: P2-11.4, P2-12.1

brute-force search¶

• Meaning: Brute-force search compares a query vector directly with every stored vector to find the nearest candidates. It uses no shortcut; it checks the whole candidate set.
• Why it matters: For small data, brute-force search is simple and exact. As the vector collection grows, it shows why search indexes and approximate nearest-neighbor methods become necessary: the number of comparisons grows with the number of stored candidates.
• Related concepts: search index, ANN, approximate nearest neighbor, similarity search
• Core Section: P1-13.4
• Appears in: P1-13.2