Explainable AI — Free Practice Questions

Free explainable AI practice questions, every answer explained. Covers SHAP, LIME, interpretability and model transparency.

30 questions · answers explained · free to practise

▶️ Start the interactive quiz

Topics covered

Sample questions with answers

8 of the 30 questions in this set, with the correct answer marked and every option explained.

1. What is the primary goal of Explainable AI (XAI)?

  • To reduce the size of neural networks
    That is model compression — pruning, quantisation, distillation. A smaller model is not automatically an understandable one.
  • To improve model accuracy
    Explainability and accuracy are largely independent, and there is often a trade-off: the most accurate models tend to be the least transparent. XAI aims to make a model's reasoning legible, not to raise its score.
  • To increase the amount of training data
    A data collection concern. More data may improve performance but reveals nothing about how the model decides.
  • To make machine learning models transparent and interpretable
    The goal of Explainable AI is to make the behavior of machine learning models transparent and interpretable to humans, allowing for a better understanding of how decisions are made.

2. Which of the following is a commonly used post-hoc explainability technique?

  • LIME (Local Interpretable Model-agnostic Explanations)
    LIME is a popular post-hoc method used to explain the predictions of machine learning models by approximating the model locally with an interpretable one.
  • Early Stopping
    A regularisation technique that halts training when validation performance plateaus. It affects how the model is trained, not how its decisions are explained.
  • Batch Normalization
    Normalises activations between layers to stabilise training. An architectural component with no interpretive function.
  • Gradient Descent
    The optimisation algorithm that fits the model. Gradients are used by some explanation methods, but descent itself explains nothing.

3. Which of the following models is considered inherently interpretable?

  • Support Vector Machine
    A linear SVM is reasonably interpretable through its coefficients, but with a non-linear kernel the decision boundary lives in a transformed space that cannot be read directly.
  • Deep Neural Network
    Millions of weights spread across many layers, with no human-readable decision path. The canonical black box.
  • Decision Tree
    Decision trees are considered inherently interpretable because their decision-making process can be easily visualized and understood by following the paths from the root to the leaf nodes.
  • Random Forest
    Each individual tree is readable, but a forest averages hundreds of them. That ensembling is exactly what buys accuracy and costs interpretability.

4. What does SHAP (SHapley Additive exPlanations) aim to provide in the context of Explainable AI?

  • Data augmentation strategies
    Augmentation expands a training set with transformed examples. It concerns the data pipeline, not attribution of a prediction.
  • Regularization techniques to reduce overfitting
    Regularisation constrains the model during training. SHAP is applied after training to a model already fitted.
  • Fair distribution of feature contributions to the model's output
    SHAP values provide a way to fairly distribute contributions among the features to the model's prediction, based on Shapley values from cooperative game theory.
  • Optimal hyperparameter tuning
    Hyperparameter search finds better settings. SHAP explains the behaviour of whatever model those settings produced.

5. Which type of feature importance measures how much a model's accuracy drops when a particular feature is shuffled?

  • Permutation Importance
    Permutation importance measures the importance of a feature by randomly shuffling its values and evaluating how much the model's performance deteriorates.
  • Gradient Clipping
    Caps gradient magnitude during training to prevent instability. A training-time safeguard, not a measure of feature importance.
  • Weight Initialization
    Sets starting values before training begins. It influences convergence, not which features matter afterwards.
  • L1 Regularization
    Drives some coefficients to exactly zero, which does perform a kind of feature selection. But it acts during training by shaping the model, rather than measuring importance in a model already trained.

6. Which of the following techniques provides local interpretability?

  • SHAP values
    SHAP values provide local interpretability by explaining individual predictions, while techniques like PDP offer global insights into how features influence predictions.
  • Support Vector Machine
    A model, not an explanation technique. A linear SVM is somewhat interpretable, but it is the thing being explained.
  • Gradient Boosting
    Also a model — an ensemble of sequentially fitted trees. It supplies built-in importance scores, but those are global rankings rather than explanations of one prediction.
  • Partial Dependence Plot (PDP)
    An explanation technique, but a global one. It shows the average effect of a feature across the whole dataset, which by construction cannot explain why one particular instance received its prediction.

7. How does Explainable AI help in mitigating bias in machine learning models?

  • By ensuring all features have equal importance
    Forcing equal importance would destroy the model's predictive power, and equal weighting is not the same as fairness. Some features legitimately matter more.
  • By increasing model complexity
    Added complexity makes a model harder to scrutinise. Bias detection depends on being able to see what the model is doing.
  • By generating synthetic data
    Synthetic or rebalanced data is a mitigation technique in its own right, but it is not explainability. XAI's contribution is exposing that a problem exists in the first place.
  • By providing transparency into decision-making and identifying biased features
    Explainable AI helps identify biased features and model decisions, allowing practitioners to take corrective actions and ensure fairness in model predictions.

8. Which technique can be used to explain the output of a black-box model by approximating it with an interpretable model?

  • LIME
    LIME (Local Interpretable Model-agnostic Explanations) approximates the predictions of a black-box model locally using a simpler and more interpretable surrogate model.
  • Backpropagation
    Computes gradients so weights can be updated. Part of training the black box, not of interpreting it.
  • Gradient Boosting
    An ensemble modelling technique. It builds a predictor rather than approximating an existing one for explanatory purposes.
  • Dropout
    Randomly deactivates units during training to combat overfitting. A regulariser with no interpretive role.

22 more questions in the app

Practise the full 30-question set with a timer, scoring and progress tracking.

Start the free quiz
Get the ad-free PRO app

More practice sets

Browse every quiz, tutorial and interactive AI tool on the All Tutorials & Tools page, or jump to a certification hub: AWS, Azure AI, Google Cloud, AWS Data Engineer.