Advanced Machine Learning (AML) Certification — Questions and Answers

Question 1: What is the purpose of max pooling in a convolutional neural network?

Reduce spatial dimensions while retaining dominant features (Correct answer)
Introduce non-linearity into the network
Normalize activations across the batch
Increase the spatial resolution of feature maps

Correct answer: Reduce spatial dimensions while retaining dominant features

Max pooling downsamples feature maps by taking the maximum value in each pooling window, reducing size while preserving strong activations.

Question 2: In AML practice, what is a needs assessment?

An employee benefits survey
A real estate appraisal
A list of office supply requirements
A systematic process to identify gaps between current conditions and desired outcomes (Correct answer)

Correct answer: A systematic process to identify gaps between current conditions and desired outcomes

A needs assessment systematically identifies and evaluates the gaps between current performance or conditions and desired outcomes, providing data-driven justification for programs, training, or interventions.

Question 3: Which outlier detection method uses the interquartile range (IQR) to flag extreme values?

Min-max scaling
Z-score normalization
Winsorization
Tukey's fence method (Correct answer)

Correct answer: Tukey's fence method

Tukey's fence method flags values beyond 1.5x IQR from Q1 or Q3 as outliers for removal or treatment.

Question 4: Under the EU AI Act, which risk category requires mandatory conformity assessments before deployment?

Limited risk
High risk (Correct answer)
Minimal risk
Prohibited risk

Correct answer: High risk

High-risk AI systems under the EU AI Act must undergo mandatory conformity assessments and meet strict requirements before market deployment.

Question 5: Which architecture is most commonly used for image-to-image tasks like semantic segmentation and medical image analysis?

ResNet
U-Net (encoder-decoder with skip connections) (Correct answer)
VGG-16
Inception v3

Correct answer: U-Net (encoder-decoder with skip connections)

U-Net uses a contracting encoder path and an expanding decoder path with skip connections that preserve spatial detail for precise pixel prediction.

Question 6: What does TF-IDF stand for in text feature extraction?

Token Filter–Index Document Format
Term Frequency–Inverse Document Frequency (Correct answer)
Text Format–Inverse Data Frame
Text Frequency–Inverse Document Function

Correct answer: Term Frequency–Inverse Document Frequency

TF-IDF weighs each word by how often it appears in a document (TF) penalized by how common it is across all documents (IDF).

Question 7: Which algorithm is commonly used in supervised learning?

Autoencoders
Linear Regression (Correct answer)
PCA
K-Means Clustering

Correct answer: Linear Regression

Linear Regression is a classic and widely used algorithm in supervised learning, specifically for regression tasks. It models the relationship between a dependent variable and one or more independent variables by fitting a linear equation to the observed data. This algorithm requires labeled data where the output variable is continuous, making it a prime example of a supervised approach.

Question 8: The concept of 'human-in-the-loop' is mandated for high-risk AI systems under the EU AI Act primarily to:

Satisfy labor union requirements for workforce participation
Reduce computational costs of inference
Ensure meaningful human oversight and the ability to intervene or override AI decisions (Correct answer)
Meet ISO 9001 quality management standards

Correct answer: Ensure meaningful human oversight and the ability to intervene or override AI decisions

The EU AI Act requires high-risk systems to allow human oversight so that operators can understand, monitor, and intervene in AI decisions to prevent harm.

Question 9: Which metric is typically used to evaluate classification algorithms?

Mean Squared Error
R-squared
Silhouette Score
Accuracy (Correct answer)

Correct answer: Accuracy

Accuracy is a widely used and intuitive metric for evaluating classification algorithms. It measures the proportion of correctly predicted instances (both true positives and true negatives) out of the total number of instances in the dataset. While other metrics like precision, recall, and F1-score provide more nuanced insights, accuracy offers a straightforward overall measure of a classifier's performance.

Question 10: When reporting model drift in production, which type of drift refers to changes in the input feature distribution?

Covariate shift (Correct answer)
Label drift
Prior probability shift
Concept drift

Correct answer: Covariate shift

Covariate shift occurs when the marginal distribution of input features P(X) changes while the conditional P(Y|X) remains stable.

Question 11: Which technique generates synthetic minority class samples to address class imbalance?

Stratified sampling
Bootstrapping
Random undersampling
SMOTE (Correct answer)

Correct answer: SMOTE

SMOTE creates new synthetic samples by interpolating between existing minority class examples rather than duplicating them.

Question 12: In MLflow, what is the purpose of the 'Model Registry' component?

To schedule automated retraining pipelines
To manage model versioning, staging, and production transitions (Correct answer)
To store raw training datasets
To visualize training metrics in real time

Correct answer: To manage model versioning, staging, and production transitions

MLflow Model Registry provides a centralized hub for managing the full lifecycle of ML models including versioning and stage transitions (Staging, Production, Archived).

Question 13: Which of the following BEST describes 'dual-use' risk in the context of machine learning research?

Training a model on both labeled and unlabeled data simultaneously
Research that can be applied for both beneficial and harmful purposes (Correct answer)
Using two different ML frameworks for the same task
Sharing a model across two separate business units

Correct answer: Research that can be applied for both beneficial and harmful purposes

Dual-use risk refers to technologies or research that can be intentionally or unintentionally repurposed to cause harm.

Question 14: In a typical CI/CD pipeline for ML (MLOps), what step immediately follows model training and precedes deployment?

Feature engineering
Infrastructure provisioning
Model evaluation and validation (Correct answer)
Data ingestion

Correct answer: Model evaluation and validation

After training, the model must pass evaluation gates (accuracy thresholds, fairness checks, performance benchmarks) before it is approved for deployment.

Question 15: Which evaluation metric for generative language models measures how well a probability model predicts a sample by computing the exponential of average negative log-likelihood?

BLEU
Perplexity (Correct answer)
ROC-AUC
F1 Score

Correct answer: Perplexity

Perplexity quantifies how surprised a language model is by new text; lower perplexity indicates better predictive performance.

Question 16: What is a machine learning data pipeline?

A visualization tool for feature distributions
A method for hyperparameter tuning
An automated sequence of data processing steps from ingestion to model-ready format (Correct answer)
A storage system for training datasets

Correct answer: An automated sequence of data processing steps from ingestion to model-ready format

A data pipeline automates and chains data collection, cleaning, transformation, and feature engineering steps for reproducibility.

Question 17: Which automated retraining trigger strategy is considered a best practice in production MLOps?

Scheduled or event-driven retraining triggered by performance degradation thresholds or new data availability (Correct answer)
Retraining only when a new model architecture is designed
Manual retraining only on a fixed monthly schedule
Retraining after every new user request to the API

Correct answer: Scheduled or event-driven retraining triggered by performance degradation thresholds or new data availability

Automated retraining pipelines that trigger on both schedules and performance thresholds keep models current without requiring manual intervention.

Question 18: Which method identifies predictive features by measuring how much each reduces impurity in a trained Random Forest?

Variance Inflation Factor (VIF)
Recursive Feature Elimination
Feature importance from Random Forest (Correct answer)
Correlation matrix analysis

Correct answer: Feature importance from Random Forest

Random Forest models compute feature importance scores based on average impurity reduction contributed by each feature across all trees.

Question 19: What is the primary purpose of the Expected Improvement (EI) acquisition function in Bayesian optimization?

To ensemble predictions from multiple surrogates
To balance exploration and exploitation by quantifying improvement probability over the current best (Correct answer)
To directly minimize the surrogate model's prediction
To sample uniformly from unexplored regions

Correct answer: To balance exploration and exploitation by quantifying improvement probability over the current best

EI computes the expected value of improvement over the current best observation, naturally trading off exploration (uncertainty) against exploitation (predicted value).

Question 20: What is the importance of continuing education in AML professional practice?

It is only required for new professionals
To justify higher fees
To maintain current knowledge, meet certification requirements, and adapt to evolving industry standards (Correct answer)
To fill time between client appointments

Correct answer: To maintain current knowledge, meet certification requirements, and adapt to evolving industry standards

Continuing education ensures professionals stay current with new developments, technologies, and best practices in their field while fulfilling certification renewal requirements and providing better service to clients.

Question 21: What is A/B testing in the context of ML model deployment?

Testing models on two separate servers simultaneously
Running automated unit tests on model training code
Comparing two model versions by routing live traffic to each and measuring performance differences (Correct answer)
Testing two different datasets against the same model

Correct answer: Comparing two model versions by routing live traffic to each and measuring performance differences

A/B testing routes a portion of real production traffic to a challenger model while the rest serves the incumbent, enabling data-driven model selection.

Question 22: What distinguishes a Restricted Boltzmann Machine (RBM) from a standard autoencoder as unsupervised representation learners?

RBMs learn deterministic encodings; autoencoders learn probabilistic latent representations
RBMs are generative probabilistic models; autoencoders are deterministic encoder-decoder architectures (Correct answer)
RBMs use backpropagation; autoencoders use contrastive divergence
RBMs require labeled data; autoencoders operate unsupervised

Correct answer: RBMs are generative probabilistic models; autoencoders are deterministic encoder-decoder architectures

RBMs are energy-based generative models trained with contrastive divergence that learn a probability distribution over inputs, while autoencoders deterministically encode inputs to reconstruct them.

Question 23: In AML practice, what is the purpose of a standard operating procedure (SOP)?

To satisfy management preferences only
To create unnecessary paperwork
To restrict employee creativity
To document step-by-step instructions for routine tasks to ensure consistency and quality (Correct answer)

Correct answer: To document step-by-step instructions for routine tasks to ensure consistency and quality

SOPs provide standardized, detailed instructions for routine operations, ensuring consistency, quality, efficiency, and safety regardless of which qualified individual performs the task.

Question 24: In TensorFlow Serving, which protocol is preferred for high-throughput, low-latency model inference in production?

GraphQL
WebSockets
REST/HTTP
gRPC (Correct answer)

Correct answer: gRPC

gRPC uses Protocol Buffers and HTTP/2 for binary serialization and multiplexing, providing lower latency and higher throughput than REST for TF Serving inference requests.

Question 25: Why is transparency important in AI systems?

It helps users trust AI decisions (Correct answer)
It prevents data sharing.
It reduces computational costs.
It hides proprietary algorithms.

Correct answer: It helps users trust AI decisions

Transparency in AI systems is crucial because it allows users to understand how and why an AI model arrived at a particular decision or prediction. When the decision-making process is clear and interpretable, it builds trust and confidence in the AI's outputs. This is especially important in sensitive applications where accountability and user acceptance are paramount.

Question 26: What fundamentally distinguishes SARSA from Q-learning?

SARSA targets continuous action spaces while Q-learning targets discrete ones
SARSA requires an environment model while Q-learning does not
SARSA uses neural networks while Q-learning uses lookup tables
SARSA is on-policy and updates using the actual next action taken; Q-learning is off-policy and updates using the greedy max-Q action (Correct answer)

Correct answer: SARSA is on-policy and updates using the actual next action taken; Q-learning is off-policy and updates using the greedy max-Q action

SARSA (on-policy) updates Q-values using the action actually taken by the behavior policy, while Q-learning (off-policy) always bootstraps from the greedy maximum Q-value action.

Question 27: What does professional liability insurance protect a AML practitioner against?

Financial loss from claims of negligence, errors, or omissions in professional services (Correct answer)
Natural disaster damage to the office
Employee health expenses
Theft of office equipment

Correct answer: Financial loss from claims of negligence, errors, or omissions in professional services

Professional liability insurance (errors and omissions coverage) protects practitioners from the financial consequences of claims alleging negligence, mistakes, or failure to perform professional duties, covering legal defense costs and settlements.

Question 28: What is the credit assignment problem in reinforcement learning?

Distributing shared rewards fairly among agents in cooperative settings
Determining how to allocate compute between training and inference
Assigning numerical credit scores to competing reward function designs
Identifying which past actions were responsible for a reward received much later in a sequence (Correct answer)

Correct answer: Identifying which past actions were responsible for a reward received much later in a sequence

The credit assignment problem is the difficulty of determining which earlier actions in a long trajectory caused a delayed reward, requiring techniques like eligibility traces or TD(λ) to propagate credit backward.

Question 29: What is the role of an ML model registry in the MLOps lifecycle?

To serve as a central catalog for tracking, versioning, and managing trained model artifacts through the deployment lifecycle (Correct answer)
To automate hyperparameter optimization across training runs
To monitor live model predictions and alert on drift
To store raw training datasets and feature CSVs

Correct answer: To serve as a central catalog for tracking, versioning, and managing trained model artifacts through the deployment lifecycle

A model registry provides a governed repository where teams can register, version, stage, and promote models from development to production.

Question 30: In the context of Neural Architecture Search (NAS), what does the 'weight sharing' strategy in one-shot NAS achieve?

It trains multiple architectures sequentially to find the best
It constrains the search to architectures with identical parameter counts
It reuses pretrained weights from ImageNet for all candidate architectures
It trains a single supernet where all candidate architectures share parameters, dramatically reducing search cost (Correct answer)

Correct answer: It trains a single supernet where all candidate architectures share parameters, dramatically reducing search cost

One-shot NAS trains a supernet once where all sub-architectures share parameters, then evaluates candidates by sampling from the supernet, reducing search cost from thousands of GPU-days to hours.

Question 31: What is semantic segmentation in computer vision?

Ranking image regions by confidence score
Generating a text caption for an image
Assigning a class label to every pixel in an image (Correct answer)
Detecting and localizing objects with bounding boxes

Correct answer: Assigning a class label to every pixel in an image

Semantic segmentation classifies each pixel of an image into a category, producing a pixel-wise class map of the entire scene.

Question 32: Which NLP technique converts words into dense vector representations that capture semantic relationships?

Bag of Words
Stemming
TF-IDF
Word2Vec / Word Embeddings (Correct answer)

Correct answer: Word2Vec / Word Embeddings

Word embeddings like Word2Vec map words to dense vectors where semantically similar words have high cosine similarity.

Question 33: What is the purpose of peer review in AML professional practice?

To create competition between colleagues
To determine who should be promoted
To reduce workload through delegation
To evaluate work quality through assessment by qualified colleagues and promote continuous improvement (Correct answer)

Correct answer: To evaluate work quality through assessment by qualified colleagues and promote continuous improvement

Peer review provides objective quality assessment by qualified professionals, identifying areas for improvement, validating practices, and promoting professional accountability and continuous quality enhancement.

Question 34: What should a AML professional do when they discover an ethical violation by a colleague?

Report through appropriate organizational channels while documenting the observed behavior (Correct answer)
Post about it on social media
Ignore it to maintain workplace harmony
Confront the colleague publicly

Correct answer: Report through appropriate organizational channels while documenting the observed behavior

Professionals have an ethical obligation to report violations through proper channels (supervisor, ethics committee, licensing board). Documentation of observations ensures accuracy and supports the investigation process.

Question 35: What is the purpose of model versioning in MLOps?

To label different datasets used to train a model
To increment model complexity with each training run
To compress model weights for faster inference
To track distinct model iterations with metadata to enable rollback and comparison (Correct answer)

Correct answer: To track distinct model iterations with metadata to enable rollback and comparison

Model versioning logs each trained model artifact with its configuration, metrics, and data lineage to support governance and rollback.

Question 36: What is the purpose of a train-validation-test split in model development?

To tune hyperparameters and evaluate final model performance on unseen data (Correct answer)
To increase the size of the training set
To balance class distributions
To perform feature engineering

Correct answer: To tune hyperparameters and evaluate final model performance on unseen data

The validation set enables hyperparameter tuning while the test set provides an unbiased final performance estimate.

Question 37: What is shadow deployment in machine learning operations?

Deploying a model to a private server without public access
Running a model only during off-peak hours to save resources
Running a new model in parallel with the production model without using its predictions for actual decisions (Correct answer)
Testing a model with synthetic data before live deployment

Correct answer: Running a new model in parallel with the production model without using its predictions for actual decisions

Shadow deployment routes live traffic to both the current and new model, logging new model outputs for evaluation without affecting end users.

Question 38: In AML practice, what is evidence-based practice?

Using only the newest methods regardless of evidence
Integrating the best available research evidence with professional expertise and client needs (Correct answer)
Following only personal experience and intuition
Doing whatever the client requests

Correct answer: Integrating the best available research evidence with professional expertise and client needs

Evidence-based practice combines rigorous research evidence, professional clinical expertise, and client/patient preferences and values to make informed decisions that optimize outcomes.

Question 39: What is the effect of using a very small batch size in stochastic gradient descent?

Smoother gradient estimates with faster convergence
Elimination of the need for learning rate scheduling
Reduced memory usage with no effect on convergence quality
Higher gradient noise that can escape local minima but slower wall-clock convergence (Correct answer)

Correct answer: Higher gradient noise that can escape local minima but slower wall-clock convergence

Small batch sizes produce noisy gradient estimates that can help escape sharp local minima but typically require more iterations and wall-clock time to converge.

Question 40: What is the purpose of subword tokenization methods like Byte-Pair Encoding (BPE) used in models like GPT?

Convert text to binary representations for efficient storage
Replace all punctuation tokens before model training
Speed up inference by reducing vocabulary size to single characters
Balance vocabulary size and out-of-vocabulary handling by splitting rare words into frequent subword units (Correct answer)

Correct answer: Balance vocabulary size and out-of-vocabulary handling by splitting rare words into frequent subword units

BPE iteratively merges frequent character pairs into subword units, allowing models to handle rare and unseen words by decomposing them.

Question 41: Which technique addresses the vanishing gradient problem by allowing gradients to flow directly through skip connections?

Dropout regularization
Residual connections (ResNets) (Correct answer)
Batch normalization alone
Weight decay

Correct answer: Residual connections (ResNets)

Residual connections let gradients bypass layers via identity shortcuts, preventing them from vanishing in very deep networks.

Question 42: In AML practice, what is the purpose of a standard operating procedure (SOP)?

To create unnecessary paperwork
To restrict employee creativity
To satisfy management preferences only
To document step-by-step instructions for routine tasks to ensure consistency and quality (Correct answer)

Correct answer: To document step-by-step instructions for routine tasks to ensure consistency and quality

SOPs provide standardized, detailed instructions for routine operations, ensuring consistency, quality, efficiency, and safety regardless of which qualified individual performs the task.

Question 43: What is the primary function of a model monitoring dashboard in production?

To automate model retraining on a fixed schedule
To display hyperparameter search results from training runs
To track production metrics like prediction drift, latency, and data quality in real time (Correct answer)
To visualize the model's internal architecture and weights

Correct answer: To track production metrics like prediction drift, latency, and data quality in real time

Monitoring dashboards give operations teams real-time visibility into model health, enabling rapid detection of and response to performance degradation.

Question 44: What does model quantization achieve during ML model deployment?

It reduces model size and speeds up inference by using lower-precision numerical representations (Correct answer)
It encrypts model weights for secure deployment in regulated environments
It converts regression models into classification models automatically
It increases model accuracy by adding more learnable parameters

Correct answer: It reduces model size and speeds up inference by using lower-precision numerical representations

Quantization converts 32-bit floating point weights to 8-bit integers, drastically reducing model size and inference latency with minimal accuracy loss.

Question 45: What is the primary advantage of hierarchical reinforcement learning (HRL) for complex tasks?

It guarantees convergence to the global optimum in non-convex reward landscapes
It decomposes complex long-horizon tasks into sub-goals, improving sample efficiency and enabling sub-behavior reuse (Correct answer)
It enables parallel training across multiple processors with no communication overhead
It eliminates the need for reward engineering by automatically deriving intrinsic reward signals

Correct answer: It decomposes complex long-horizon tasks into sub-goals, improving sample efficiency and enabling sub-behavior reuse

HRL breaks long-horizon tasks into hierarchies where high-level policies set sub-goals for lower-level policies, dramatically improving sample efficiency and allowing learned sub-behaviors to transfer across tasks.

Question 46: When a stakeholder misrepresents model capabilities in an external presentation, what is the most appropriate response?

Publicly contradict the stakeholder during the presentation
Ignore the misrepresentation to avoid conflict
Allow the misrepresentation to stand to protect organizational reputation
Privately correct the stakeholder with accurate information and offer to help prepare factually accurate materials for future communications (Correct answer)

Correct answer: Privately correct the stakeholder with accurate information and offer to help prepare factually accurate materials for future communications

Private correction paired with an offer to provide accurate materials addresses the immediate issue diplomatically while preventing future misrepresentations.

Question 47: Which technique is used to augment training images by randomly flipping, rotating, or cropping them?

Data augmentation (Correct answer)
Weight decay
Dropout regularization
Batch normalization

Correct answer: Data augmentation

Data augmentation artificially expands the training set with label-preserving transformations, reducing overfitting in vision models.

Question 48: What does professional liability insurance protect a AML practitioner against?

Natural disaster damage to the office
Employee health expenses
Financial loss from claims of negligence, errors, or omissions in professional services (Correct answer)
Theft of office equipment

Correct answer: Financial loss from claims of negligence, errors, or omissions in professional services

Question 49: In the context of ML risk management, 'operational risk' most directly refers to:

The risk of choosing the wrong evaluation metric
Overfitting caused by insufficient regularization
Failures in people, processes, or systems supporting the model's operation (Correct answer)
The chance that the chosen loss function is suboptimal

Correct answer: Failures in people, processes, or systems supporting the model's operation

Operational risk in ML covers failures stemming from inadequate internal processes, human error, system failures, or external events affecting the model's production pipeline.

Question 50: Which technique handles missing values by replacing them with the average of the column?

Mean imputation (Correct answer)
One-hot encoding
Standardization
Label encoding

Correct answer: Mean imputation

Mean imputation replaces missing values with the column's average, preserving the dataset's overall distribution.

Question 51: Which statistical test is recommended for comparing multiple ML models across multiple datasets to avoid inflated Type I error?

Paired t-test with Bonferroni correction
Friedman test followed by post-hoc Nemenyi test (Correct answer)
ANOVA with Tukey's HSD
Wilcoxon signed-rank test

Correct answer: Friedman test followed by post-hoc Nemenyi test

The Friedman test is a non-parametric rank-based test for multiple classifiers over multiple datasets, and the Nemenyi post-hoc test identifies which pairs differ significantly.

Question 52: What does CI/CD stand for in an MLOps pipeline?

Continuous Improvement / Continuous Diagnostics
Containerized Inference / Continuous Development
Continuous Integration / Continuous Deployment (Correct answer)
Central Intelligence / Cloud Deployment

Correct answer: Continuous Integration / Continuous Deployment

CI/CD automates testing and deployment pipelines so new model versions can be validated and released quickly and reliably.

Question 53: Which loss function is most appropriate for training a neural network on a multi-label classification task where each sample can belong to multiple classes?

Binary cross-entropy applied independently per label (Correct answer)
Categorical cross-entropy with softmax
Mean squared error
Hinge loss with multi-class margin

Correct answer: Binary cross-entropy applied independently per label

Binary cross-entropy treats each label independently with a sigmoid activation, allowing multiple classes to be simultaneously positive.

Question 54: Which feature engineering technique creates new features from products or powers of existing features?

Polynomial feature expansion (Correct answer)
Imputation
Feature scaling
Label encoding

Correct answer: Polynomial feature expansion

Polynomial feature expansion generates new features as products or powers of existing ones, enabling linear models to capture non-linear patterns.

Question 55: Which statement about t-SNE (t-Distributed Stochastic Neighbor Embedding) is TRUE?

t-SNE is computationally efficient and scales well to millions of data points
t-SNE is deterministic and always produces the same embedding
t-SNE preserves global structure better than local structure
t-SNE uses Student's t-distribution in low-dimensional space to alleviate the crowding problem (Correct answer)

Correct answer: t-SNE uses Student's t-distribution in low-dimensional space to alleviate the crowding problem

t-SNE uses a heavy-tailed t-distribution in the low-dimensional map, which allows moderately dissimilar points to be placed further apart and alleviates the crowding problem.

Question 56: Which dimensionality reduction technique projects data onto directions of maximum variance?

UMAP
LDA
PCA (Correct answer)
t-SNE

Correct answer: PCA

PCA (Principal Component Analysis) finds orthogonal axes of maximum variance to compress feature dimensions.

Question 57: What is the primary risk of using a static holdout test set for ongoing model quality monitoring over multiple retraining cycles?

Overfitting to the validation set through hyperparameter tuning
Test set contamination through repeated model selection on the same data (Correct answer)
Increased inference cost per evaluation cycle
Instability in precision-recall tradeoffs

Correct answer: Test set contamination through repeated model selection on the same data

Repeated selection of models based on the same test set causes the test set to implicitly influence model development, leaking information and inflating performance estimates.

Question 58: Which metric category is most critical when monitoring a deployed classification model in production?

Training loss from the most recent training run
Model file size on disk
Prediction confidence distribution and accuracy on live data (Correct answer)
Number of API calls per second

Correct answer: Prediction confidence distribution and accuracy on live data

Monitoring live prediction confidence and accuracy reveals model drift and performance degradation before it significantly impacts business outcomes.

Question 59: What is Inverse Reinforcement Learning (IRL) primarily designed to accomplish?

Training agents through episodes played in reverse chronological order
Inferring an unknown reward function from observed expert demonstrations (Correct answer)
Learning policies that minimize rather than maximize cumulative reward
Training agents to reverse or undo previously taken actions

Correct answer: Inferring an unknown reward function from observed expert demonstrations

IRL infers the underlying reward function from expert behavioral demonstrations, which is useful when the reward function is unknown, hard to define, or too complex to specify manually.

Question 60: In NLP, what does Named Entity Recognition (NER) identify in text?

Duplicate sentences across a corpus
Real-world entities such as people, organizations, and locations (Correct answer)
The overall sentiment of a document
The grammatical role of each word in a sentence

Correct answer: Real-world entities such as people, organizations, and locations

NER tags spans of text that refer to specific entities like person names, companies, dates, and geographic locations.

Question 61: What is the primary purpose of model explainability tools like SHAP in production ML systems?

To automate feature engineering pipelines
To compress model artifacts for faster deployment
To speed up model inference at scale
To explain individual predictions and ensure model transparency for stakeholders and auditors (Correct answer)

Correct answer: To explain individual predictions and ensure model transparency for stakeholders and auditors

SHAP attributes each feature's contribution to individual predictions, supporting transparency and regulatory compliance in production.

Question 62: What is the key distinction between model-based and model-free reinforcement learning?

Model-free RL requires more labeled data than model-based approaches
Model-based RL only works in discrete action spaces
Model-based RL uses deep neural networks; model-free uses tabular methods only
Model-based RL learns or uses an environment model for planning; model-free learns directly from interactions (Correct answer)

Correct answer: Model-based RL learns or uses an environment model for planning; model-free learns directly from interactions

Model-based RL builds or leverages an explicit model of environment dynamics for planning, while model-free RL learns policies or value functions directly from experience without modeling transitions.

Question 63: What does a log transformation primarily help with during feature engineering?

Encoding categorical variables
Reducing the skewness of right-skewed distributions (Correct answer)
Handling missing data
Normalizing binary features

Correct answer: Reducing the skewness of right-skewed distributions

Log transformation compresses large values and expands small ones, making right-skewed distributions more symmetric.

Question 64: Which tool is commonly used to containerize ML models for consistent deployment across environments?

NumPy
Jupyter Notebook
Docker (Correct answer)
Scikit-learn

Correct answer: Docker

Docker packages ML models with their dependencies into portable containers that run consistently across development and production environments.

Question 65: What is the primary benefit of feature selection in a machine learning pipeline?

Reduces overfitting and training time (Correct answer)
Adds synthetic training data
Increases model complexity
Improves data imputation accuracy

Correct answer: Reduces overfitting and training time

Feature selection removes irrelevant or redundant features, reducing overfitting risk and computational cost.

Question 66: An ML practitioner is hired to consult on a project but lacks expertise in the specific domain. According to professional ethics standards, the MOST appropriate response is to:

Accept the project and learn on the job to avoid revenue loss
Deliver results without mentioning the expertise gap
Decline or disclose limitations and bring in a domain expert (Correct answer)
Charge a lower fee to compensate for reduced quality

Correct answer: Decline or disclose limitations and bring in a domain expert

Professional ethics require practitioners to practice only within their competence and to be transparent about limitations.

Question 67: What is 'data versioning' in the context of ML tools like DVC (Data Version Control)?

Encrypting sensitive training data before storage
Compressing datasets to reduce storage costs
Splitting datasets into versioned train/validation/test splits
Tracking changes to large datasets and model artifacts using Git-like version control (Correct answer)

Correct answer: Tracking changes to large datasets and model artifacts using Git-like version control

DVC adds Git-like version control for large data files and ML artifacts by storing metadata in Git while the actual data lives in remote storage (S3, GCS, etc.).

Question 68: What is training-serving skew in machine learning deployment?

The accuracy gap between training and test set performance
The mismatch between model size and available hardware capacity
The time lag between completing model training and completing deployment
Discrepancies between how features are computed during offline training versus online model serving (Correct answer)

Correct answer: Discrepancies between how features are computed during offline training versus online model serving

Training-serving skew occurs when feature computation logic differs between offline training and online serving, causing unexpected prediction errors in production.

Question 69: What is model drift?

The gradual improvement of a model over time
A regularization technique for production models
The movement of model weights during training
The degradation of model performance as real-world data distribution changes from training data (Correct answer)

Correct answer: The degradation of model performance as real-world data distribution changes from training data

Model drift occurs when the statistical properties of input data or target relationships change post-deployment, reducing model accuracy.

Question 70: When publishing an ML research paper, a researcher finds that a competing team's unpublished work led them to the same solution. What is the ethically required action?

Cite the competing team's work if it influenced the research (Correct answer)
Contact the journal to claim priority based on submission date
Publish independently since the competing work is unpublished
Delay publication until the competing team publishes first

Correct answer: Cite the competing team's work if it influenced the research

Academic integrity requires attributing any work that influenced your research, regardless of its publication status.

Question 71: Which NLP preprocessing step reduces words to their base or root form by removing suffixes (e.g., 'running' → 'run')?

Lemmatization
POS tagging
Dependency parsing
Stemming (Correct answer)

Correct answer: Stemming

Stemming applies rule-based suffix stripping to reduce words to an approximate root, which may not be a valid dictionary word.

Question 72: What is MLOps in the context of advanced machine learning?

The practice of combining ML development with operations to deploy and maintain models in production (Correct answer)
A type of neural network architecture
A dataset versioning tool
A programming language for machine learning

Correct answer: The practice of combining ML development with operations to deploy and maintain models in production

MLOps applies DevOps principles to machine learning, automating training, deployment, monitoring, and retraining workflows.

Question 73: Equalized odds as a fairness criterion requires that a classifier satisfies:

Equal precision across groups
Equal positive prediction rates across groups
Equal negative predictive value across groups
Equal true positive and false positive rates across groups (Correct answer)

Correct answer: Equal true positive and false positive rates across groups

Equalized odds, defined by Hardt et al., requires both equal true positive rates (TPR) and equal false positive rates (FPR) across protected groups simultaneously.

Question 74: What is a REST API in the context of machine learning model deployment?

A monitoring dashboard for deployed ML models
A standardized interface that allows applications to send data to a model and receive predictions via HTTP (Correct answer)
A database for storing model artifacts
A version control system for datasets

Correct answer: A standardized interface that allows applications to send data to a model and receive predictions via HTTP

A REST API exposes model inference as HTTP endpoints so any downstream application can request predictions via standard web requests.

Question 75: Which stakeholder management approach is best when dealing with a highly technical stakeholder who wants to micromanage model architecture decisions?

Block all communication from this stakeholder
Exclude them from technical meetings entirely
Agree with all their suggestions regardless of merit
Establish a structured technical review process with defined decision rights and consultation points (Correct answer)

Correct answer: Establish a structured technical review process with defined decision rights and consultation points

A structured review process with defined decision rights respects technical expertise while maintaining project governance and team autonomy.

Question 76: In cohort analysis, what does retention rate measure?

The fraction of users from an initial cohort who remain active in a later period (Correct answer)
The churn rate of the bottom decile
The average revenue per user cohort
The percentage of new users acquired in a period

Correct answer: The fraction of users from an initial cohort who remain active in a later period

Retention rate tracks what fraction of users who joined in a given cohort period are still active N periods later.

Question 77: Which transformer-based model introduced bidirectional context for language understanding and set new NLP benchmarks in 2018?

GPT-2
BERT (Correct answer)
ELMo
Seq2Seq

Correct answer: BERT

BERT (Bidirectional Encoder Representations from Transformers) pre-trains on masked language modeling using full left and right context simultaneously.

Question 78: What role does a 'kill switch' or 'circuit breaker' play in ML model risk management?

It terminates training early when validation loss plateaus
It disables GPU acceleration during high-traffic periods
It removes outlier training examples before each epoch
It automatically halts or reverts model predictions when performance drops below a defined threshold (Correct answer)

Correct answer: It automatically halts or reverts model predictions when performance drops below a defined threshold

A kill switch or circuit breaker provides an automated safeguard that takes the model offline or reverts to a fallback rule-based system when real-time metrics signal unacceptable degradation.

Question 79: Why is k-fold cross-validation used when evaluating feature engineering choices?

To increase the training dataset size
To automatically select the optimal number of features
To handle class imbalance
To obtain a more reliable performance estimate across multiple data subsets (Correct answer)

Correct answer: To obtain a more reliable performance estimate across multiple data subsets

K-fold cross-validation rotates the validation set across k subsets, giving a more robust estimate than a single train-test split.

Question 80: What is the primary advantage of using Isolation Forest over LOF for anomaly detection?

Isolation Forest is supervised and needs labeled anomalies
Isolation Forest scales efficiently to high-dimensional, large datasets (Correct answer)
Isolation Forest produces calibrated anomaly probabilities
Isolation Forest requires no hyperparameters

Correct answer: Isolation Forest scales efficiently to high-dimensional, large datasets

Isolation Forest has linear time complexity and scales well to high dimensions and large datasets by isolating anomalies with random partitioning, unlike the quadratic-complexity LOF.

Question 81: What is the key difference between normalization and standardization?

They are identical techniques with different names
Normalization encodes categories; standardization scales numerics
Normalization removes outliers; standardization does not
Normalization scales values to [0,1]; standardization transforms data to zero mean and unit variance (Correct answer)

Correct answer: Normalization scales values to [0,1]; standardization transforms data to zero mean and unit variance

Normalization maps values to a fixed range like [0,1] while standardization transforms data to have mean 0 and standard deviation 1.

Question 82: What role does documentation play in AML regulatory compliance?

It is only needed for tax purposes
It creates unnecessary paperwork
It is optional for experienced professionals
It provides evidence of compliance, supports audits, and creates a defensible record of professional activities (Correct answer)

Correct answer: It provides evidence of compliance, supports audits, and creates a defensible record of professional activities

Documentation is the foundation of compliance, providing verifiable evidence that regulatory requirements have been met, supporting audit processes, and creating a defensible record if compliance is ever questioned.

Question 83: In Bayesian hyperparameter optimization, what probabilistic model is most commonly used to approximate the objective function?

Neural network meta-learner
Support vector regressor
Gaussian Process (Correct answer)
Random forest surrogate

Correct answer: Gaussian Process

Gaussian Processes provide a posterior distribution over the objective function, enabling uncertainty-aware acquisition functions like Expected Improvement.

Question 84: What is canary deployment in ML production systems?

Testing new models in a completely isolated validation environment
Using a lightweight fallback model when the primary model fails
Gradually rolling out a new model to a small subset of users before a full release (Correct answer)
Deploying a model only for internal QA testing

Correct answer: Gradually rolling out a new model to a small subset of users before a full release

Canary deployment incrementally increases traffic to a new model version, allowing real-world validation with minimal risk exposure.

Question 85: Which artifact BEST serves as the single source of truth for tracking ML experiment reproducibility across a team?

A README file updated manually by each team member
Slack messages with attached screenshots of training curves
An MLflow or similar experiment tracking registry logging code version, data hash, params, and metrics (Correct answer)
A shared Google Doc with hyperparameter tables

Correct answer: An MLflow or similar experiment tracking registry logging code version, data hash, params, and metrics

Experiment tracking platforms like MLflow capture code version, data lineage, hyperparameters, and metrics in a queryable, reproducible format.

Question 86: What is target encoding for categorical variables?

Encoding the target variable as integers
Replacing each category value with the mean of the target variable for that category (Correct answer)
Encoding text features using TF-IDF
Binarizing the target for classification tasks

Correct answer: Replacing each category value with the mean of the target variable for that category

Target encoding substitutes each category with the average target value for that category, directly capturing predictive power.

Question 87: A cross-functional team disagrees on model deployment timelines due to differing risk tolerances. What communication strategy should the ML lead use?

Let each team implement their preferred timeline independently
Force a technical decision without input from other teams
Delay deployment indefinitely to avoid conflict
Facilitate a structured risk assessment discussion using a shared framework to align all stakeholders (Correct answer)

Correct answer: Facilitate a structured risk assessment discussion using a shared framework to align all stakeholders

A shared risk assessment framework creates common ground and objective criteria that help cross-functional teams reach consensus on deployment timing.

Question 88: In NLP, what is tokenization?

Removing stop words from a sentence
Splitting raw text into individual units such as words or subwords (Correct answer)
Converting words to their base grammatical form
Mapping tokens to their embedding vectors

Correct answer: Splitting raw text into individual units such as words or subwords

Tokenization breaks raw text into discrete units (tokens) such as words, subwords, or characters that a model can process.

Question 89: What is 'tail risk' in the context of ML model predictions, and why is it critical for risk management?

The risk of rare but extreme prediction errors that can cause disproportionately large negative outcomes (Correct answer)
The risk that the last few layers of a neural network overfit
The risk of slow inference for long-tail categorical features
The risk that evaluation metrics are computed on the wrong data split

Correct answer: The risk of rare but extreme prediction errors that can cause disproportionately large negative outcomes

Tail risk refers to low-probability but high-severity errors at the extremes of the prediction distribution, which standard average-based metrics obscure but which can dominate real-world losses.

Question 90: What is transfer learning in the context of computer vision?

Transferring labels from one dataset to another
Using a pre-trained model's weights as a starting point for a new task (Correct answer)
Training a model from scratch on a new dataset
Copying model architecture without pre-trained weights

Correct answer: Using a pre-trained model's weights as a starting point for a new task

Transfer learning reuses the feature representations learned from a large dataset (e.g., ImageNet) to improve performance on a smaller target task.

Question 91: What is the primary function of 'gradient checkpointing' in deep learning?

Saving model checkpoints when validation loss improves
Trading compute for memory by recomputing activations during backpropagation instead of storing them (Correct answer)
Checkpointing distributed training progress for fault tolerance
Clipping gradients to prevent exploding gradient issues

Correct answer: Trading compute for memory by recomputing activations during backpropagation instead of storing them

Gradient checkpointing discards intermediate activations during the forward pass and recomputes them during backpropagation, reducing peak memory usage at the cost of ~33% more computation.

Question 92: What is 'concept drift' and why is it a risk management concern in deployed ML systems?

Random variation in model outputs due to floating-point precision
Instability in gradient descent caused by poor initialization
Gradual memory leak in model serving infrastructure
A change in the statistical relationship between inputs and the target variable over time (Correct answer)

Correct answer: A change in the statistical relationship between inputs and the target variable over time

Concept drift occurs when the underlying data-generating process changes so that previously learned input-output relationships no longer hold, degrading model reliability.

Question 93: What is the primary purpose of feature scaling in machine learning?

To encode categorical variables
To remove outliers from the dataset
To ensure all features contribute equally to model training (Correct answer)
To reduce the number of features

Correct answer: To ensure all features contribute equally to model training

Feature scaling normalizes feature ranges so that no single feature dominates model training due to its magnitude.

Question 94: Which encoding technique assigns integer values to categories that have a natural order?

Ordinal encoding (Correct answer)
Binary encoding
Frequency encoding
One-hot encoding

Correct answer: Ordinal encoding

Ordinal encoding maps ordered categories such as low/medium/high to integers that preserve their natural rank.

Question 95: Which NLP task involves assigning a label (positive, negative, neutral) to a piece of text based on its emotional tone?

Coreference Resolution
Named Entity Recognition
Sentiment Analysis (Correct answer)
Machine Translation

Correct answer: Sentiment Analysis

Sentiment analysis classifies text according to the opinion or emotion expressed, commonly as positive, negative, or neutral.

Question 96: What is the curse of dimensionality in machine learning?

The difficulty of visualizing high-dimensional data
The phenomenon where data becomes sparse as dimensions increase, degrading model performance (Correct answer)
The problem of having too many training samples
The computational cost of training deep neural networks

Correct answer: The phenomenon where data becomes sparse as dimensions increase, degrading model performance

As feature dimensions increase, data points become increasingly sparse, making distance-based algorithms less effective.

Question 97: What does the Variance Inflation Factor (VIF) measure in regression analysis?

The importance of features in a regression model
The variance of individual feature distributions
The degree of multicollinearity among predictor variables (Correct answer)
The inflation of model error as more features are added

Correct answer: The degree of multicollinearity among predictor variables

VIF quantifies how much the variance of a regression coefficient is inflated due to correlation with other predictor variables.

Question 98: Which cross-validation strategy is most appropriate when your dataset has significant temporal ordering?

Shuffle-split CV
Leave-one-out CV
Time-series split (walk-forward) (Correct answer)
Stratified k-fold

Correct answer: Time-series split (walk-forward)

Time-series split (walk-forward validation) prevents data leakage by always training on past data and validating on future data.

Advanced Machine Learning (AML) Certification

The AML certification validates advanced competency across the full machine learning lifecycle, covering core algorithms, feature engineering, model deployment, NLP, computer vision, and professional skills including AI ethics, governance, and stakeholder communication.

100

Questions

120 min

Time Limit

70%

To Pass

Checking for saved progress...

Exam Rules

You can skip questions and return to them later
Flag questions for review before submitting
No feedback shown until you submit the entire exam
Unanswered questions count as wrong — answer everything
10 pretest questions are mixed in and don't affect your score
Timer auto-submits when time runs out
Your progress is auto-saved every 30 seconds