I85PREP
⌘ K
Huawei HCIE-AI study materials
Huawei HCIE-AI study resources

Huawei HCIE-AI practice questions

One question at a time. Build your Huawei HCIE-AI understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1AI fundamentals and machine learning

In supervised learning for AI model training, what is the primary role of the labeled dataset?

  1. It provides input-output pairs used to compute the loss and update model parameters Correct answer.
  2. It stores the final deployed model weights for inference
  3. It supplies unlabeled samples for the model to cluster by similarity
  4. It defines the hardware topology used during distributed training

Explanation

Supervised learning requires ground-truth labels so the model can compare its prediction with the target and compute a loss. Gradients derived from that loss drive parameter updates. Unlabeled data and infrastructure are separate concerns.

Question 2AI fundamentals and machine learning

A team notices a model performs very well on training data but poorly on unseen validation data. Which term best describes this behavior?

  1. Gradient vanishing
  2. Underfitting
  3. Data leakage
  4. Overfitting Correct answer.

Explanation

Overfitting occurs when a model memorizes training-specific noise and fails to generalize to new samples. The large gap between training and validation performance is its classic symptom. Underfitting instead shows poor performance on both sets.

Question 3AI fundamentals and machine learning

Which activation function outputs values in the range from zero to one and is commonly used in binary classification output layers?

  1. Sigmoid Correct answer.
  2. Tanh
  3. Softmax
  4. ReLU

Explanation

The sigmoid function maps any real input to the interval (0, 1), which can be interpreted as a probability for binary classification. Softmax is used for multi-class outputs, and ReLU and Tanh have different ranges.

Question 4AI fundamentals and machine learning

In gradient descent, what does the learning rate control?

  1. The number of layers in the network
  2. The proportion of data held out for validation
  3. The number of classes the model can predict
  4. The size of the parameter update applied at each step Correct answer.

Explanation

The learning rate scales the gradient before it is applied to the parameters, directly determining step size. A value too large can cause divergence, while too small slows convergence. It does not affect architecture or data splits.

Question 5AI fundamentals and machine learning

Which metric is most appropriate for evaluating a classifier on a highly imbalanced dataset where the positive class is rare?

  1. Mean squared error
  2. Raw accuracy
  3. Area under the precision-recall curve Correct answer.
  4. R-squared

Explanation

Accuracy can look high simply by predicting the majority class, hiding poor detection of the rare class. The precision-recall curve and its area focus on the positive class and are more informative under imbalance. MSE and R-squared are regression metrics.

Question 6AI fundamentals and machine learning

A model trained on data collected in one region is deployed in another region and produces degraded predictions. Which concept does this illustrate?

  1. Distribution shift Correct answer.
  2. Dropout
  3. Batch normalization
  4. Regularization

Explanation

Distribution shift occurs when the deployment data distribution differs from the training distribution, degrading performance. Regularization, batch normalization, and dropout are training techniques that do not describe this mismatch.

Question 7AI fundamentals and machine learning

What is the main purpose of splitting data into training, validation, and test sets?

  1. To reduce the total amount of data needed
  2. To increase the number of features available
  3. To guarantee the model cannot overfit
  4. To tune hyperparameters on validation data and estimate final performance on unseen test data Correct answer.

Explanation

The validation set guides hyperparameter selection, while the test set provides an unbiased estimate of generalization performance. Keeping them separate prevents optimistic bias. Splitting does not reduce data needs or add features.

Question 8AI fundamentals and machine learning

Which technique applies L1 regularization to a linear model?

  1. Adding the sum of squared weights to the loss
  2. Increasing the learning rate
  3. Removing the bias term from the model
  4. Adding the sum of absolute weights to the loss Correct answer.

Explanation

L1 regularization adds the sum of absolute parameter values to the loss, which tends to produce sparse weights. L2 adds the sum of squared weights. Neither involves removing bias or changing the learning rate.

Question 9AI fundamentals and machine learning

A neural network for image classification uses convolutional layers. What property do these layers primarily exploit?

  1. Sequential dependencies between words
  2. Global ordering of pixel values
  3. Local spatial patterns and weight sharing Correct answer.
  4. Exact reproduction of input dimensions

Explanation

Convolutional layers apply shared filters over local receptive fields, capturing spatial patterns like edges and textures while reducing parameter count. This weight sharing is central to their efficiency on images. Sequential word dependencies are handled by sequence models.

Question 10AI fundamentals and machine learning

What does the softmax function produce in a multi-class classification network's output layer?

  1. A one-hot encoding of the input
  2. A probability distribution over the classes that sums to one Correct answer.
  3. A set of unbounded real-valued logits
  4. A single binary probability

Explanation

Softmax exponentiates logits and normalizes them so the outputs form a probability distribution summing to one. This supports multi-class cross-entropy training. A single binary probability comes from sigmoid, and raw logits are pre-softmax values.

Question 11AI fundamentals and machine learning

Which optimizer maintains per-parameter adaptive learning rates using estimates of first and second moments of the gradients?

  1. Stochastic gradient descent with momentum
  2. Grid search
  3. Adam Correct answer.
  4. Newton's method

Explanation

Adam tracks exponential moving averages of the gradient and its square, giving each parameter an adaptive step size. Momentum-only SGD uses just first-moment-like accumulation. Newton's method and grid search are not adaptive-moment optimizers.

Question 12AI fundamentals and machine learning

A data scientist wants to reduce overfitting in a deep network without collecting more data. Which combination of techniques is most directly appropriate?

  1. Increase model depth and remove regularization
  2. Raise the learning rate and remove validation
  3. Add dropout and weight decay Correct answer.
  4. Increase batch size and remove early stopping

Explanation

Dropout and weight decay are regularization techniques that constrain the model's ability to memorize training noise. Both are standard remedies for overfitting when additional data is unavailable. Increasing capacity or removing safeguards would worsen the problem.

Question 13AI fundamentals and machine learning

Which statement correctly describes the bias-variance tradeoff?

  1. High variance indicates the model is too rigid to capture the underlying pattern
  2. High bias corresponds to underfitting, and high variance corresponds to overfitting Correct answer.
  3. High bias indicates excessive sensitivity to training noise
  4. Bias and variance can both be eliminated by adding more layers

Explanation

High bias means the model is too simple and underfits, while high variance means it is too sensitive to training data and overfits. The tradeoff describes balancing these two error sources. Adding layers alone does not eliminate both.

Question 14AI fundamentals and machine learning

What is the effect of using a very large learning rate during gradient descent training?

  1. The training data is automatically augmented
  2. The loss may oscillate or diverge instead of converging Correct answer.
  3. The model is guaranteed to reach the global minimum faster
  4. The number of parameters decreases automatically

Explanation

An excessively large learning rate causes updates to overshoot minima, producing oscillation or divergence. It does not guarantee faster convergence or alter model size or data. Convergence requires a suitably scaled step size.

Question 15AI fundamentals and machine learning

In k-fold cross-validation, what is the primary benefit compared with a single train-validation split?

  1. It eliminates the need for a test set
  2. It uses each sample for both training and validation across folds, giving a more stable performance estimate Correct answer.
  3. It reduces the number of training epochs required
  4. It guarantees the model will not overfit

Explanation

Cross-validation rotates the validation fold so every sample is validated once, averaging results for a lower-variance estimate. It still requires a separate test set for final evaluation and does not guarantee against overfitting. Epoch count is unrelated.

Question 16AI fundamentals and machine learning

Which loss function is standard for training a multi-class classifier with mutually exclusive classes?

  1. Hinge loss for regression
  2. Kullback-Leibler divergence between input features
  3. Categorical cross-entropy Correct answer.
  4. Mean absolute error

Explanation

Categorical cross-entropy measures the divergence between predicted class probabilities and one-hot targets, matching softmax outputs. MAE is a regression loss. Hinge loss is used for margin-based classification, and KL divergence between features is not a standard classification objective.

Question 17AI fundamentals and machine learning

A recommendation system must learn from user-item interactions without any explicit labels. Which learning paradigm fits best?

  1. Supervised regression
  2. Linear discriminant analysis
  3. Unsupervised or self-supervised representation learning Correct answer.
  4. Reinforcement learning with a fixed reward

Explanation

Without explicit labels, the system can learn structure from interaction patterns using unsupervised or self-supervised objectives such as predicting co-occurrence. Supervised regression and discriminant analysis require labeled targets. Reinforcement learning needs a defined reward signal.

Question 18AI fundamentals and machine learning

What does the term 'epoch' mean in neural network training?

  1. One forward and backward pass over a single sample
  2. The total number of parameters in the model
  3. One complete pass through the entire training dataset Correct answer.
  4. The size of the validation set

Explanation

An epoch is one full traversal of the training dataset, typically composed of many mini-batch updates. A single sample pass is an iteration or step. Parameter count and validation size are unrelated quantities.

Question 19AI fundamentals and machine learning

Which preprocessing step is most important before feeding features with very different scales into a gradient-based model?

  1. Feature standardization or normalization Correct answer.
  2. Removing the target variable
  3. One-hot encoding all continuous features
  4. Increasing the number of classes

Explanation

Features on very different scales can cause ill-conditioned gradients and slow or unstable convergence, so standardization or normalization is standard practice. One-hot encoding continuous features loses information. Target removal and class count changes are unrelated.

Question 20AI fundamentals and machine learning

In a transformer architecture, what is the function of the self-attention mechanism?

  1. It reduces the number of training epochs
  2. It performs convolutional downsampling of images
  3. It computes pairwise relevance between positions to aggregate contextual information Correct answer.
  4. It replaces the need for any training data

Explanation

Self-attention computes attention scores between all positions, allowing each token to aggregate information weighted by relevance. This captures long-range dependencies without recurrence. It does not affect epoch count, data requirements, or perform convolutional downsampling.

Question 21AI fundamentals and machine learning

A team observes that adding more training data no longer improves validation performance and the model has high training accuracy. Which conclusion is most reasonable?

  1. The model is likely overfitting and needs stronger regularization or better data quality Correct answer.
  2. The test set was mislabeled
  3. The learning rate is too low to matter
  4. The model is underfitting and needs more capacity

Explanation

High training accuracy with plateaued validation performance and diminishing returns from more data suggests the model has memorized the training distribution. Stronger regularization or improved data quality is appropriate. Underfitting would show low training accuracy.

Question 22AI fundamentals and machine learning

Which statement about precision and recall is correct?

  1. Precision and recall are always equal for binary classifiers
  2. Recall measures the fraction of predicted positives that are truly positive
  3. Precision measures the fraction of predicted positives that are correct, and recall measures the fraction of actual positives that were retrieved Correct answer.
  4. Precision measures the fraction of actual positives that were retrieved

Explanation

Precision is true positives divided by predicted positives, while recall is true positives divided by actual positives. The two definitions are distinct and often trade off. They are equal only in special cases, not generally.

Question 23AI fundamentals and machine learning

What is the primary purpose of batch normalization in deep network training?

  1. To perform data augmentation during inference
  2. To reduce internal covariate shift and stabilize optimization by normalizing layer inputs Correct answer.
  3. To replace the need for activation functions
  4. To increase the number of learnable parameters without bound

Explanation

Batch normalization normalizes activations across a mini-batch, which stabilizes and accelerates optimization and can act as a mild regularizer. It introduces learnable scale and shift terms but does not replace activations. Data augmentation is unrelated to inference normalization.

Question 24AI fundamentals and machine learning

A model achieves 99% accuracy on a dataset where 99% of samples belong to one class. What is the most accurate interpretation?

  1. Accuracy is the best metric regardless of class balance
  2. The model has definitely learned all class boundaries
  3. The model may simply be predicting the majority class and is likely ineffective at detecting the minority class Correct answer.
  4. The model is excellent because accuracy is high

Explanation

When one class dominates, predicting it always yields accuracy equal to its prevalence, so 99% accuracy can reflect no real learning. Metrics such as recall, precision, and F1 for the minority class are needed. High accuracy alone is uninformative under severe imbalance.

Question 25AI fundamentals and machine learning

Which factor most directly determines the receptive field size of a neuron in a deep convolutional network?

  1. The learning rate schedule
  2. The number of training epochs
  3. The size of the batch used for inference
  4. The kernel sizes and stride of preceding layers Correct answer.

Explanation

Receptive field grows with the kernel sizes and strides of the layers preceding a neuron, along with depth. Training epochs, learning rate schedule, and inference batch size do not change the architecture's receptive field. This is an architectural property.

Question 26AI fundamentals and machine learning

In an AI development pipeline, which activity belongs specifically to feature engineering rather than to model training?

  1. Tuning the number of layers in a residual network
  2. Minimizing a cross-entropy objective with mini-batch stochastic gradient descent
  3. Converting raw timestamps into cyclical hour-of-day and day-of-week encodings Correct answer.
  4. Selecting the number of epochs through early stopping

Explanation

Feature engineering transforms raw inputs into representations that make patterns easier for a learner to capture, such as cyclical encodings of time. Training-time operations like gradient updates, architecture depth selection, and epoch control act on the model rather than on the input representation.

Question 27AI fundamentals and machine learning

A team needs to estimate how a trained model will behave on unseen production traffic before deployment. Which evaluation practice gives the most defensible estimate?

  1. Measuring inference latency on the training batch
  2. Reporting accuracy computed on the same data used for gradient updates
  3. Scoring the model on a held-out test set never used during training or tuning Correct answer.
  4. Averaging the training loss curves across the final epochs

Explanation

A held-out test set that was untouched during training and hyperparameter selection provides an unbiased estimate of generalization performance. Evaluating on training data measures fit to seen examples, and loss curves or latency measurements say nothing about accuracy on new traffic.

Question 28AI fundamentals and machine learning

Which machine learning paradigm is characterized by an agent learning a policy from reward signals produced by its interactions with an environment?

  1. Reinforcement learning Correct answer.
  2. Supervised learning
  3. Self-supervised pretraining
  4. Unsupervised learning

Explanation

Reinforcement learning trains an agent to maximize cumulative reward through trial-and-error interaction with an environment, without a fixed labeled target per input. Supervised and self-supervised methods rely on target signals derived from data, and unsupervised methods seek structure without rewards.

Question 29AI fundamentals and machine learning

A dataset contains a feature recording monthly income in a currency with values ranging from 0 to 5,000,000, alongside a feature recording years of education from 0 to 25. A gradient-boosted tree model is being trained. Which statement about scaling is most accurate?

  1. Income must be log-transformed or the model will diverge during training
  2. Tree-based splits are invariant to monotonic feature transformations, so scaling is not required Correct answer.
  3. Education must be divided by income to make the features comparable
  4. Both features must be standardized, otherwise the tree cannot split on income

Explanation

Decision-tree models choose split thresholds by ordering feature values, so any monotonic transformation such as standardization or min-max scaling leaves the possible splits and resulting trees unchanged. Gradient-based neural models are the ones that typically require comparable feature scales.

Question 30AI fundamentals and machine learning

Which statement best describes the role of the learning rate in gradient-based training of a neural network?

  1. It determines the number of features retained by the input layer
  2. It sets the number of samples processed before each parameter update
  3. It controls the fraction of neurons randomly dropped during each forward pass
  4. It scales the size of the parameter step taken along the negative gradient direction Correct answer.

Explanation

The learning rate multiplies the gradient when updating parameters, directly controlling step size along the descent direction. Too large a value can cause oscillation or divergence, while too small a value slows convergence.

Question 31AI fundamentals and machine learning

A classification model for rare disease screening has recall of 0.95 and precision of 0.30. The clinical team wants fewer false alarms while keeping recall high. Which change to the decision threshold is most appropriate to try first?

  1. Raise the threshold so only higher-confidence samples are predicted positive Correct answer.
  2. Keep the threshold fixed and remove the positive class from training data
  3. Keep the threshold fixed and increase the number of training epochs
  4. Lower the threshold so more samples are predicted positive

Explanation

Raising the decision threshold makes the positive prediction stricter, which typically reduces false positives and therefore raises precision, at the cost of possibly lowering recall. Lowering the threshold does the opposite and would increase false alarms.

Question 32AI fundamentals and machine learning

Which description correctly captures the difference between a parameter and a hyperparameter in a trained neural network?

  1. Parameters are learned by optimization, while hyperparameters are set by the practitioner before or during training Correct answer.
  2. Parameters are chosen before training, while hyperparameters are learned from gradients
  3. Parameters control data preprocessing, while hyperparameters control deployment
  4. Parameters are always scalars, while hyperparameters are always tensors

Explanation

Model parameters such as weights and biases are adjusted by the optimization algorithm using training data. Hyperparameters such as learning rate, batch size, and layer count are specified externally and govern how training proceeds.

Question 33AI fundamentals and machine learning

An engineer trains a model on data from one hospital and finds it performs poorly at a second hospital with different scanner vendors. Which term best describes this failure mode?

  1. Label leakage
  2. Gradient vanishing
  3. Domain shift Correct answer.
  4. Underfitting

Explanation

Domain shift occurs when the distribution of inputs at deployment differs from the training distribution, as with different scanner characteristics across hospitals. The model's learned mapping may be valid but misaligned with the new data distribution.

Question 34AI fundamentals and machine learning

Which metric is most appropriate for evaluating a regression model when large errors are especially costly and should be penalized disproportionately?

  1. Accuracy
  2. Mean absolute error
  3. Root mean squared error Correct answer.
  4. Area under the ROC curve

Explanation

Root mean squared error squares residuals before averaging, so large deviations dominate the score and are penalized more heavily than under mean absolute error. Accuracy and AUC are classification metrics and do not apply to continuous targets.

Question 35AI fundamentals and machine learning

What is the primary effect of applying dropout during training of a deep neural network?

  1. It randomly deactivates units per forward pass, discouraging co-adaptation and acting as regularization Correct answer.
  2. It increases the effective learning rate for early layers
  3. It permanently removes low-magnitude weights from the network
  4. It normalizes activations across the batch to stabilize training

Explanation

Dropout randomly zeroes a fraction of activations during each training step, which prevents units from relying on specific partners and reduces overfitting. At inference, scaling is applied so the expected activation magnitude is preserved.

Question 36AI fundamentals and machine learning

A team must choose between a deep model with 200 million parameters and a linear model with 5,000 parameters for a tabular dataset of 3,000 samples. Which choice is most defensible from a bias-variance perspective?

  1. The linear model, because limited data makes the high-capacity model prone to high variance Correct answer.
  2. The deep model, because more parameters always yield higher accuracy
  3. Either model, because parameter count has no relation to generalization
  4. The deep model, because linear models cannot represent any useful function

Explanation

With only 3,000 samples, a very high-capacity model has ample freedom to memorize noise, producing high variance and poor generalization. A lower-capacity linear model is more likely to generalize unless strong regularization or far more data is available.

Question 37AI fundamentals and machine learning

In a convolutional neural network, what is the main effect of inserting a pooling layer between convolutional stages?

  1. It increases the number of learnable filters in the following layer
  2. It applies a nonlinear activation to each individual pixel
  3. It reduces the spatial resolution and introduces local translation invariance Correct answer.
  4. It normalizes the channel statistics of the previous layer

Explanation

Pooling aggregates values over local windows, shrinking spatial dimensions and making the representation less sensitive to small translations of the input. It has no learnable parameters and is distinct from activation or normalization operations.

Question 38AI fundamentals and machine learning

A model trained on a dataset with 500 positive and 9,500 negative samples reports 92% accuracy but misses most positive cases. Which approach most directly addresses this class imbalance?

  1. Increase the number of output classes
  2. Report accuracy on the training set instead of the test set
  3. Apply class weighting or resampling so positive examples contribute more to the loss Correct answer.
  4. Remove the negative class from the dataset

Explanation

Class weighting or resampling changes the effective contribution of each class to the objective, pushing the model to learn the minority class rather than defaulting to the majority. Accuracy alone is misleading under imbalance, so metrics such as recall and F1 should accompany the change.

Question 39AI fundamentals and machine learning

Which statement about the role of an activation function in a feedforward neural network is correct?

  1. It replaces the need for a loss function during training
  2. It ensures all gradients remain constant across layers
  3. It introduces nonlinearity so stacked layers can represent non-linear mappings Correct answer.
  4. It guarantees that the loss surface is convex

Explanation

Without a nonlinear activation, a stack of linear layers collapses into a single linear transformation. Nonlinear activations such as ReLU or tanh enable the network to approximate complex functions.

Question 40AI fundamentals and machine learning

An engineer notices that a deep network's gradients shrink to near zero in early layers during backpropagation. Which architectural change most directly mitigates this?

  1. Increasing the batch size while keeping depth unchanged
  2. Adding residual skip connections that carry information across layers Correct answer.
  3. Replacing the loss function with mean absolute error
  4. Removing all normalization layers from the network

Explanation

Residual connections provide shortcut paths for gradients to flow backward, which alleviates vanishing gradients in deep stacks. Batch size, loss choice, and normalization removal do not create such direct gradient pathways.

Question 41AI fundamentals and machine learning

Which statement best explains why an ensemble of decision trees trained with bagging usually generalizes better than a single deep decision tree?

  1. The ensemble uses a different loss function than the single tree
  2. Each tree in the ensemble is trained on the full dataset with no randomization
  3. Bagging increases the bias of each individual tree to zero
  4. Averaging many trees reduces variance while preserving low bias Correct answer.

Explanation

Bagging trains trees on bootstrap samples and aggregates their predictions, so individual errors partially cancel, lowering variance. Each tree retains relatively low bias, which is why the combined model often generalizes better than one deep tree.

Question 42AI fundamentals and machine learning

What does the softmax function output in a multi-class classification network?

  1. A set of scores summing to one that can be interpreted as class probabilities Correct answer.
  2. A binary label for each class independently
  3. A normalized feature embedding for the input
  4. A single scalar representing the predicted class index

Explanation

Softmax exponentiates logits and normalizes them so the outputs are non-negative and sum to one, giving a probability distribution over mutually exclusive classes. It is typically paired with cross-entropy loss during training.

Question 43AI fundamentals and machine learning

A team is deciding how to handle missing values in a numeric feature that is missing for 30% of samples. Which consideration is most important when choosing between mean imputation and adding a missing-indicator feature?

  1. Missingness may itself carry predictive signal, so encoding it can preserve information Correct answer.
  2. Mean imputation always yields better accuracy than any alternative
  3. Adding an indicator feature guarantees the model will overfit
  4. Missing values must be deleted because models cannot accept imputed inputs

Explanation

If the fact that a value is missing correlates with the target, an indicator variable lets the model exploit that signal, whereas mean imputation alone erases it. The choice should be validated empirically rather than assumed.

Question 44AI fundamentals and machine learning

Which statement accurately describes the purpose of a validation set during model development?

  1. It provides the final unbiased estimate reported to stakeholders
  2. It replaces the need for a test set in all cases
  3. It guides hyperparameter and model selection while remaining separate from training Correct answer.
  4. It is used to compute gradients for parameter updates

Explanation

The validation set is used to compare models and tune hyperparameters, so it influences development decisions. The test set, kept untouched until the end, provides the final unbiased estimate.

Question 45AI fundamentals and machine learning

A language model is pretrained to predict masked tokens in unlabeled text and then fine-tuned on labeled sentiment data. Which description of this pipeline is correct?

  1. Both stages are unsupervised because the text is unlabeled
  2. Both stages require human-annotated labels for every token
  3. The pretraining stage is reinforcement learning and fine-tuning is supervised
  4. The pretraining stage is self-supervised and the fine-tuning stage is supervised Correct answer.

Explanation

Masked-token prediction constructs its own targets from unlabeled text, making it self-supervised. Fine-tuning on sentiment labels uses human-provided targets, which is supervised learning.

Question 46AI fundamentals and machine learning

Which factor most directly determines the receptive field of a neuron in a deep convolutional network?

  1. The number of output channels in the final layer
  2. The number of samples in each training batch
  3. The kernel sizes and strides of the preceding layers Correct answer.
  4. The value of the learning rate used during training

Explanation

Receptive field grows with the spatial extent covered by preceding kernels and is affected by strides and dilations, which determine how far input pixels influence a given unit. Channel counts, learning rate, and batch size do not change the spatial mapping.

Question 47AI fundamentals and machine learning

A practitioner wants to reduce variance in a k-nearest-neighbors classifier on a noisy dataset. Which change is most likely to achieve this?

  1. Switch the distance metric to a random assignment
  2. Decrease k so fewer neighbors vote
  3. Remove feature scaling from the pipeline
  4. Increase k so each prediction averages over more neighbors Correct answer.

Explanation

Increasing k smooths the decision boundary by averaging over more neighbors, which typically lowers variance at the cost of some bias. Decreasing k makes predictions depend on fewer points and increases variance.

Question 48AI fundamentals and machine learning

An organization deploys a model whose predictions influence loan decisions. Which practice best supports responsible AI in this context?

  1. Retraining continuously without any evaluation of deployed behavior
  2. Maximizing overall accuracy while ignoring differences between demographic groups
  3. Monitoring subgroup error rates and documenting the model's limitations and intended use Correct answer.
  4. Removing all documentation to avoid liability exposure

Explanation

Responsible deployment requires measuring performance across relevant subgroups and documenting limitations, assumptions, and intended use so that risks are visible to operators and auditors. Aggregate accuracy alone can hide harms to specific groups.

Question 49AI fundamentals and machine learning

A model's training loss keeps decreasing while its validation loss begins to rise after epoch 12. Which interpretation and action are most appropriate?

  1. The validation set is corrupted; ignore validation loss entirely
  2. The model is overfitting after epoch 12; apply early stopping or stronger regularization Correct answer.
  3. The learning rate is too small; increase it sharply
  4. The model is underfitting; train for many more epochs

Explanation

Divergence between falling training loss and rising validation loss is the classic signature of overfitting, where the model starts fitting noise. Early stopping at the minimum validation loss or adding regularization are standard remedies.

Question 50AI fundamentals and machine learning

A team wants to compare two classifiers on the same imbalanced test set and needs a threshold-independent view of ranking quality. Which evaluation approach fits best?

  1. Plot the precision-recall curve and compute its area Correct answer.
  2. Report training time per epoch
  3. Count the number of parameters in each model
  4. Report accuracy at a single fixed threshold of 0.5

Explanation

A precision-recall curve shows performance across all thresholds, and its area summarizes ranking quality in a way that is informative under class imbalance. A single-threshold accuracy figure hides threshold sensitivity and can be dominated by the majority class.