ДалееСледующее руководство
Байесовская vs частотная статистика
Основы
РУКОВОДСТВО ПО ОСНОВАМ
Probability describes uncertainty under stated assumptions, while statistics uses observations to estimate and evaluate patterns.
Прослушайте это руководствоОзвучено ИИ · ~3 min
For ML work, the practical skill is choosing the right denominator, checking how data were sampled and explaining what a result does not establish.
Start by distinguishing an observed quantity from the population or process you want to understand. A sample mean describes the records collected, but a biased sampling process can make it misleading for a wider population. Inspect missing observations, repeated entities and how examples were selected. More rows do not automatically repair selection bias or make dependent observations independent. Learn distributions and summaries together. For the illustrative values [0, 0, 0, 20], the mean is 5 while the median is 0. Neither number is inherently wrong; each answers a different question about the data. Examine spread and unusual values before presenting one average as a complete description. Conditional probability changes the group under consideration. The fraction of true events among alerts is different from the fraction of all true events that were alerted. Suppose a toy dataset contains 1,000 cases, including 10 true events. A system flags all 10 events and 90 other cases. Its alert precision is 10 divided by 100, or 10%; its recall is 10 divided by 10, or 100%. These invented counts show why high recall can coexist with many false alerts. Uncertainty estimates depend on assumptions and the evaluation design. In the usual frequentist interpretation, a 95% confidence procedure covers a fixed population parameter in 95% of repeated samples under its assumptions; it does not promise that every resulting interval contains the truth. Keep final test data separate from repeated tuning and inspect important groups as well as averages. Prediction also differs from causation: an association alone does not show what would happen if someone changed an input.
An analyst reports both the mean of [0, 0, 0, 20], which is 5, and its median, which is 0, to show how a large value affects the summary.
A fraud team finds 10 true events among 100 alerts and reports 10% alert precision rather than confusing it with the fraction of events detected.
An evaluator keeps repeated records from the same person together when constructing a holdout split.
A researcher states the assumptions behind an interval estimate instead of treating one observed interval as a guarantee.
Evaluation tools may make uncertainty intervals and subgroup summaries easier to generate, but automatic output will not select the right population or sampling design by itself. More accessible diagnostics could expose missing groups or unstable estimates if teams preserve the relevant metadata. Practitioners will still need to explain denominators, distinguish prediction from intervention and connect errors with real consequences. As workflows change, keep the evaluation question explicit and revisit assumptions about independence and representativeness. Statistical literacy is useful for deciding which claims the evidence supports, rather than making every score look more precise.
Бесплатная рассылка
Три проверенных AI-истории каждое утро буднего дня, написанные простым английским. Бесплатно навсегда, без рекламы.
Одно электронное письмо каждый будний день. Отписаться в один клик. Мы никогда не продаем и не передаем ваш адрес.
Проверьте себя
Мгновенная обратная связь по каждому ответу и общий сертификат с проверяемым идентификатором после прохождения курса.
Поддержите бесплатное образование в области искусственного интеллекта. AI Understanding — это некоммерческая организация 501(c)(3). Никакой рекламы и платного доступа. Сделать пожертвование
Probability describes uncertainty under stated assumptions, while statistics uses observations to estimate and evaluate patterns. For ML work, the practical skill is choosing the right denominator, checking how data were sampled and explaining what a result does not establish.
The sum is 20 across four values, giving mean 5; the two middle values are both zero.
Precision is true positive alerts divided by all alerts: 10/100=10%.
Recall is detected true events divided by all true events: 10/10=100%.
Increasing sample size does not by itself remove selection bias.
Related records may leak entity-specific information across the split.
Продолжайте учиться
Другие руководства, выбранные по этой теме
ДалееСледующее руководство
Байесовская vs частотная статистика
Основы