Викторина, связанная с руководством · Жесткий Уровень
Викторина LIME «Местные пояснения»
Поймите локальный суррогат LIME, варианты возмущений и взвешивания, проверки стабильности и пределы интерпретации вклада функций.
Вопрос 1 из 8
Что LIME подгоняет вокруг выбранного предсказания?
Предварительный просмотр каждого вопроса8 вопросы
Ответы и пояснения появляются по мере прохождения теста.
- Что LIME подгоняет вокруг выбранного предсказания?
- A replacement model guaranteed to match globally
- A causal model of the observed real-world outcome
- A simple interpretable surrogate approximating the original model locally
- A single global feature ranking shared by all examples
- Почему близкие возмущения получают больший вес в локальном объяснении?
- To make every perturbed feature equally common
- To focus the surrogate on behavior near the instance
- To guarantee perturbations match the true data distribution
- To give distant decision regions equal influence on the local fit
- Объяснение LIME резко меняется при изменении семени выборки. Что должен сделать аналитик?
- Report only the explanation with the clearest feature signs
- Assume the original predictor's outputs must also change between runs
- Treat sampling variation as proof of a causal relationship
- Assess stability under reasonable settings before relying on it
- Почему независимое удаление слов может создать плохое текстовое соседство?
- The resulting sentences may be unnatural
- The original word order is guaranteed to remain semantically equivalent
- Independent removals guarantee samples from the true language distribution
- Local weighting automatically restores missing grammatical relationships
- Что устанавливает вклад признака в LIME?
- The real-world cause of the outcome
- Global feature importance for every example
- A local association in a surrogate explanation, not causation
- That a prediction is correct
- Which diagnostic directly checks whether a local surrogate represents the original model near an instance?
- Measure the predictor's global training accuracy alone
- Measure surrogate fidelity in the weighted neighborhood
- Count only the number of selected surrogate features
- Inspect only the surrogate coefficient signs
- For an image explanation, what can change which regions appear important?
- Only the display order of identical highlighted regions
- Only the text used to name an unchanged region
- Only the image file's name when pixels stay fixed
- The segmentation used as interpretable units
- Which conclusion is appropriately limited after inspecting one LIME explanation?
- This surrogate approximates this model near this selected input under stated settings
- The model uses features identically for all cases
- The highlighted factor caused the real outcome
- The whole model is fair and trustworthy