Working glossary and notation

Working glossary and notation — table 1
Term	Meaning and small example
Observation	One unit being studied: a wine sample, message or service-day
Feature / X	Information available when making a prediction; X has rows by input columns
Target / y	Outcome to predict; one label/value per observation
Parameter	A learned number, such as a coefficient or neural weight
Hyperparameter	A chosen setting, such as tree depth or epoch budget
Training	Fit parameters and learned preprocessing using development examples
Validation	Compare choices before freezing the final method
Test	Evaluate a frozen choice; viewing it repeatedly does not create fresh evidence
Leakage	Information enters learning/selection that would be unavailable in the intended decision
Baseline	A simple relevant alternative the proposed method must be compared with
Residual	Actual minus predicted; positive means underprediction
Probability / calibration	A score in 0–1; calibration asks whether comparable scores match observed frequencies
Precision / recall	TP/(TP+FP) versus TP/(TP+FN); denominators answer different questions
Batch / epoch	A subset used in one update versus one pass through training examples
Loss / metric	Objective used for fitting versus a measure used for evaluation; they may differ
Gradient	Directional sensitivity of loss to parameters; an optimizer uses it to update weights
Tensor / shape	An array and its dimensions; (32,11) means 32 rows of 11 values
Token / embedding	A text unit's identifier versus a learned numeric vector representing it
Inference	Apply fixed parameters; it does not require an unknown target
Drift	A change needing investigation; input change alone does not establish changed outcome relationships
Tool / plan	An implemented operation and a proposed invocation; a legal plan can still misunderstand intent

Use the module's foundation page for worked arithmetic and guided construction for execution. Units and available information remain part of every definition.

Terms used in the exercises

Terms used in the exercises — table 2
Term	Meaning and example	Continue
Kernel	The Python process executing notebook cells; restart clears its variables. Course notebooks use Python (course core) or Python (course frameworks)	Notebook setup
Boolean	True or False, such as `85 > 80`	Module 1
JSON	A text format for structured values, such as `{"label":"spam"}`	Module 2
Pipeline	Ordered transformations and a model fitted together on training data	Module 5
Imputation	Replacing missing inputs using a declared rule, such as a training median	Module 5
MAE	Mean absolute error: errors −2 and 4 give MAE 3 in target units	Module 5
Scaler	A fitted transformation of numerical units; learn its statistics on training	Module 6
Threshold	Cutoff converting a score into a decision; this course uses score ≥ cutoff	Module 6
Confusion matrix	Counts of TP, FP, FN and TN comparing predictions with actual labels	Module 6
False positive / FP	A positive prediction for an actual negative; a false alert	Module 6
False negative / FN	A negative prediction for an actual positive; a missed case	Module 6
Balanced accuracy	Average of recall and specificity; gives both actual classes equal weight	Module 6
Silhouette	Compares within-cluster closeness with separation from other clusters	Module 7
ReLU	Activation `max(0,x)`; −2 becomes 0, while 3 stays 3	Module 8
Sigmoid	Maps a real number into 0–1; zero maps to 0.5	Module 8
Prompt	Instructions and context supplied to a language model	Module 10
Adapter / LoRA	A small trainable update alongside frozen base weights; LoRA uses low-rank matrices	Module 11
One-hot encoding	One indicator column per category, so categories carry no implied order; fit the categories on training data	Module 3
Cross-validation	Repeatedly fitting on part of the training data and scoring on the rest, to choose settings without touching validation or test	Module 6
Gradient boosting	An ensemble that adds small trees one at a time, each correcting the errors of the previous ones; common for tabular data	Module 6
Retrieval-augmented generation (RAG)	Answering from passages retrieved for each question; evaluate the retrieval and the decision to abstain, not only the fluency of the answer	Module 10
Recall@k / MRR	Whether a relevant passage appears in the top k results, and the average of 1 / rank of the first relevant passage	Module 10
API	A defined way for programs to call a service, such as sending JSON to a prediction endpoint and receiving a status code and result	Module 12
Wilson interval	A binomial uncertainty calculation conditional on suitable sampling assumptions; not a correction for biased data	Module 13

View source for this page on GitHub