One-hot encoding<\/span><\/p>\nAnswer: 3) Pruning<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Pruning reduces the size or computational requirements of a machine learning model by removing selected parameters, connections, neurons, or other components that contribute relatively little to the model’s predictions. The objective is to maintain useful predictive performance while reducing resource consumption. Pruning can be structured or unstructured depending on how components are removed. After pruning, the model should be evaluated because removing parameters can affect accuracy, latency, and other operational characteristics. Additional fine-tuning may sometimes be used to recover performance. Pruning is particularly useful when a trained model needs to be optimized for deployment environments with limited memory or computational resources.<\/span><\/p>\nQuestion 382. What is model quantization primarily intended to reduce?<\/b><\/h3>\n
1)<\/b> Numerical precision used to represent model values<\/span>
\n<\/span>2)<\/b> Number of training labels<\/span>
\n<\/span>3)<\/b> Number of validation datasets<\/span>
\n<\/span>4)<\/b> Number of application users<\/span><\/p>\nAnswer: 1) Numerical precision used to represent model values<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Quantization reduces the numerical precision used to represent model parameters or perform calculations. For example, a model using higher-precision floating-point values may be converted to a lower-precision representation. This can reduce memory requirements and may improve inference efficiency on compatible hardware. Quantization can be performed using different strategies, including approaches that occur after training or during a training-aware process. Because lower precision can affect numerical accuracy and model predictions, the quantized model should be evaluated against the original model. Important measurements include predictive quality, latency, throughput, memory consumption, and compatibility with the intended inference environment.<\/span><\/p>\nQuestion 383. Which AWS service is designed to provide a managed environment for storing, sharing, and discovering machine learning features?<\/b><\/h3>\n
1)<\/b> Amazon CloudWatch<\/span>
\n<\/span>2)<\/b> Amazon ECR<\/span>
\n<\/span>3)<\/b> AWS CloudTrail<\/span>
\n<\/span>4)<\/b> Amazon SageMaker Feature Store<\/span><\/p>\nAnswer: 4) Amazon SageMaker Feature Store<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Amazon SageMaker Feature Store is designed to help organizations store, manage, and retrieve machine learning features for training and inference workflows. It supports centralized feature management and can help teams reuse consistently defined features across multiple machine learning applications. Feature Store provides online and offline storage capabilities for different access patterns. The online store is useful when applications require low-latency feature retrieval, while offline storage can support training and historical analysis. Centralizing feature management can also help improve consistency between training and inference. Appropriate permissions, data quality controls, and feature definitions remain important when implementing a feature management strategy.<\/span><\/p>\nQuestion 384. What is the primary purpose of an offline feature store?<\/b><\/h3>\n
1)<\/b> To serve every prediction with the lowest possible network latency<\/span>
\n<\/span>2)<\/b> To support historical feature data for training and analysis<\/span>
\n<\/span>3)<\/b> To replace the model endpoint<\/span>
\n<\/span>4)<\/b> To automatically select the final model architecture<\/span><\/p>\nAnswer: 2) To support historical feature data for training and analysis<\/b><\/p>\n
Explanation:<\/b><\/p>\n
An offline feature store is primarily useful for retaining historical feature values that can be used for machine learning training, analysis, and other batch-oriented workflows. Historical data is important because models often need examples representing conditions that existed at different points in time. An offline store can provide access to feature records without requiring the low-latency access pattern typically associated with online inference. Historical feature data can also support point-in-time correct dataset construction when the available timestamps are handled appropriately. The offline store does not itself train models or automatically select architectures; those activities remain separate parts of the ML workflow.<\/span><\/p>\nQuestion 385. What is point-in-time correctness important for when creating training datasets?<\/b><\/h3>\n
1)<\/b> Ensuring features only use information that would have been available at the prediction time<\/span>
\n<\/span>2)<\/b> Increasing the number of future records in training data<\/span>
\n<\/span>3)<\/b> Making every feature categorical<\/span>
\n<\/span>4)<\/b> Removing all timestamps from the dataset<\/span><\/p>\nAnswer: 1) Ensuring features only use information that would have been available at the prediction time<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Point-in-time correctness helps prevent future information from being accidentally included when constructing historical training examples. For each training observation, the feature values should reflect information that was available at the time the prediction would originally have been made. If information generated later is included, the model may receive an unrealistic advantage during training. This is a form of temporal leakage that can produce overly optimistic evaluation results and poor production performance. Historical feature data should therefore be joined and filtered using appropriate timestamps. Maintaining point-in-time correctness is especially important for forecasting, risk modeling, recommendations, and other systems where data changes over time.<\/span><\/p>\nQuestion 386. Which approach is most appropriate for evaluating a model when observations have a strong chronological dependency?<\/b><\/h3>\n
1)<\/b> Randomly shuffle all records before splitting<\/span>
\n<\/span>2)<\/b> Use future observations to train the model<\/span>
\n<\/span>3)<\/b> Split data according to time order<\/span>
\n<\/span>4)<\/b> Duplicate the most recent observations<\/span><\/p>\nAnswer: 3) Split data according to time order<\/b><\/p>\n
Explanation:<\/b><\/p>\n
When observations have chronological dependencies, a time-based split is generally more appropriate than randomly shuffling the entire dataset. A chronological split trains the model using earlier observations and evaluates it on later observations, better reflecting how the model will operate in production. Random splitting can allow information from future periods to influence training while evaluating earlier periods, which may produce unrealistic results. Time-based validation is particularly relevant for forecasting, financial transactions, demand prediction, and other applications where future conditions differ from historical conditions. Additional validation periods can also be used to assess whether performance remains stable across different time windows.<\/span><\/p>\nQuestion 387. What is a group-based data split intended to prevent?<\/b><\/h3>\n
1)<\/b> Excessive model parameter storage<\/span>
\n<\/span>2)<\/b> The same entity appearing across training and evaluation datasets<\/span>
\n<\/span>3)<\/b> Conversion of numerical features into vectors<\/span>
\n<\/span>4)<\/b> Use of validation metrics<\/span><\/p>\nAnswer: 2) The same entity appearing across training and evaluation datasets<\/b><\/p>\n
Explanation:<\/b><\/p>\n
A group-based split keeps related observations from the same entity together in one dataset partition. For example, if multiple transactions belong to the same customer, randomly distributing those transactions across training and test datasets may allow the model to learn customer-specific patterns from the training records and then benefit from them when evaluating the same customer. This can make performance appear better than it would be for genuinely unseen entities. Group-based splitting assigns complete groups to training, validation, or test partitions. This approach is useful for datasets involving customers, patients, devices, households, users, or other entities with multiple observations.<\/span><\/p>\nQuestion 388. What is the purpose of stratified sampling during dataset splitting?<\/b><\/h3>\n
1)<\/b> To remove every minority-class observation<\/span>
\n<\/span>2)<\/b> To ensure every feature has identical values<\/span>
\n<\/span>3)<\/b> To increase the number of model parameters<\/span>
\n<\/span>4)<\/b> To approximately preserve class proportions across dataset partitions<\/span><\/p>\nAnswer: 4) To approximately preserve class proportions across dataset partitions<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Stratified sampling divides data while attempting to preserve the relative distribution of important groups, particularly target classes, across training, validation, and test datasets. For example, if a binary dataset contains 80% negative and 20% positive examples, stratification attempts to maintain approximately similar proportions in each partition. This can produce more representative evaluation results, especially when classes are imbalanced. Without stratification, a small validation or test dataset might contain an unusually high or low proportion of minority examples. Stratification should still be applied carefully when temporal or group dependencies exist, because preserving class proportions does not automatically address other forms of data leakage.<\/span><\/p>\nQuestion 389. What is a baseline model used for in machine learning?<\/b><\/h3>\n
1)<\/b> To provide a simple reference point for evaluating more complex models<\/span>
\n<\/span>2)<\/b> To permanently replace production models<\/span>
\n<\/span>3)<\/b> To eliminate the need for test data<\/span>
\n<\/span>4)<\/b> To increase the complexity of every model<\/span><\/p>\nAnswer: 1) To provide a simple reference point for evaluating more complex models<\/b><\/p>\n
Explanation:<\/b><\/p>\n
A baseline model provides a simple reference against which more sophisticated machine learning approaches can be evaluated. Depending on the task, a baseline might use a simple statistical method, a constant prediction, a majority class, or another straightforward approach. If a complex model cannot meaningfully outperform the baseline on appropriate evaluation data, the additional complexity may not provide sufficient benefit. Baselines are also useful during experimentation because they establish a starting point for performance comparisons. A baseline should be evaluated using the same relevant dataset and metric definitions used for candidate models so that comparisons remain meaningful and consistent.<\/span><\/p>\nQuestion 390. What is a data contract in an ML data pipeline?<\/b><\/h3>\n
1)<\/b> A document describing only cloud infrastructure costs<\/span>
\n<\/span>2)<\/b> A guarantee that model accuracy will never decline<\/span>
\n<\/span>3)<\/b> A defined agreement about expected data structure, types, and quality requirements<\/span>
\n<\/span>4)<\/b> A replacement for model monitoring<\/span><\/p>\nAnswer: 3) A defined agreement about expected data structure, types, and quality requirements<\/b><\/p>\n
Explanation:<\/b><\/p>\n
A data contract defines expectations between systems that produce and consume data. In a machine learning pipeline, it can specify requirements such as field names, data types, allowed ranges, formats, required fields, and other quality conditions. Data contracts help detect upstream changes that could break preprocessing or negatively affect model predictions. For example, an upstream application changing a numerical field into a string may violate the expected contract. Automated validation can check incoming data against these requirements before the data reaches later pipeline stages. A data contract does not replace model monitoring; rather, it provides an additional control for data consistency and quality.<\/span><\/p>\nQuestion 391. Which practice helps make machine learning experiments reproducible?<\/b><\/h3>\n
1)<\/b> Changing several variables without recording them<\/span>
\n<\/span>2)<\/b> Recording datasets, code versions, parameters, and random seeds<\/span>
\n<\/span>3)<\/b> Deleting previous experiment results<\/span>
\n<\/span>4)<\/b> Using different preprocessing logic for every experiment<\/span><\/p>\nAnswer: 2) Recording datasets, code versions, parameters, and random seeds<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Reproducibility means being able to recreate an experiment or model result using the same relevant inputs and configuration. Recording information such as dataset versions, source code versions, preprocessing logic, hyperparameters, software dependencies, model configuration, and random seeds can make reproduction easier. Without this information, it may be difficult to determine why two training runs produced different results. Reproducibility is particularly valuable when comparing experiments, investigating production issues, or promoting a model through development stages. Random seeds alone do not guarantee complete reproducibility because infrastructure, library versions, hardware, and other factors can also influence results.<\/span><\/p>\nQuestion 392. Why should preprocessing transformations be versioned?<\/b><\/h3>\n
1)<\/b> To increase the number of raw records<\/span>
\n<\/span>2)<\/b> To make model endpoints unnecessary<\/span>
\n<\/span>3)<\/b> To ensure only training data is stored<\/span>
\n<\/span>4)<\/b> To reproduce the same transformation logic during training and inference<\/span><\/p>\nAnswer: 4) To reproduce the same transformation logic during training and inference<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Preprocessing transformations can directly affect the features provided to a machine learning model. If the transformation logic used during inference differs from the logic used during training, the model may receive inputs in an unexpected representation. Versioning preprocessing code and configuration helps teams reproduce the exact transformations associated with a particular model version. It also supports debugging, rollback, and controlled model promotion. Examples include scaling rules, categorical mappings, missing-value handling, feature calculations, and text processing. Preprocessing artifacts should therefore be managed as part of the model workflow rather than treated as undocumented scripts that can change independently.<\/span><\/p>\nQuestion 393. What is the main purpose of a model registry?<\/b><\/h3>\n
1)<\/b> To manage model versions and their associated metadata and lifecycle states<\/span>
\n<\/span>2)<\/b> To store only raw training data<\/span>
\n<\/span>3)<\/b> To replace all monitoring systems<\/span>
\n<\/span>4)<\/b> To automatically guarantee model accuracy<\/span><\/p>\nAnswer: 1) To manage model versions and their associated metadata and lifecycle states<\/b><\/p>\n
Explanation:<\/b><\/p>\n
A model registry provides a structured way to organize and manage machine learning model versions. It can maintain information such as model artifacts, versions, metadata, approval states, and deployment-related information. This supports controlled promotion of models through development, testing, and production workflows. A registry can also make it easier to identify which model version is currently approved or deployed and to maintain a history of previous versions. It does not guarantee model quality or automatically replace monitoring. Evaluation metrics, validation procedures, monitoring, and governance processes remain necessary to determine whether a particular model is appropriate for deployment.<\/span><\/p>\nQuestion 394. What is model lineage primarily used to track?<\/b><\/h3>\n
1)<\/b> The physical location of every prediction request<\/span>
\n<\/span>2)<\/b> The relationship between a model and its data, code, experiments, and related artifacts<\/span>
\n<\/span>3)<\/b> The number of application users<\/span>
\n<\/span>4)<\/b> The network speed of an endpoint<\/span><\/p>\nAnswer: 2) The relationship between a model and its data, code, experiments, and related artifacts<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Model lineage tracks the relationships between a machine learning model and the assets involved in creating or deploying it. These assets may include datasets, preprocessing steps, source code, experiments, training configurations, model artifacts, and deployment versions. Strong lineage makes it easier to understand where a model came from and which inputs or processes contributed to its creation. It can support troubleshooting, reproducibility, auditing, and controlled model management. For example, if a production model produces unexpected behavior, lineage information can help identify the training dataset, code version, or preprocessing configuration associated with that model.<\/span><\/p>\nQuestion 395. What is data provenance in a machine learning workflow?<\/b><\/h3>\n
1)<\/b> The history and origin of data as it moves through processing stages<\/span>
\n<\/span>2)<\/b> The number of model layers<\/span>
\n<\/span>3)<\/b> The final prediction threshold<\/span>
\n<\/span>4)<\/b> The CPU utilization of an endpoint<\/span><\/p>\nAnswer: 3) The history and origin of data as it moves through processing stages<\/b><\/p>\n
Explanation:<\/b><\/p>\n
Data provenance describes the origin, history, and transformations applied to data throughout a workflow. It can identify where data came from, which processing steps modified it, and how it contributed to a downstream dataset or model. Provenance is useful for debugging, reproducibility, auditing, and understanding the reliability of data used in machine learning. For example, if an unexpected value appears in a training dataset, provenance information can help trace it back to its source or transformation step. Maintaining clear provenance can also support governance by making the lifecycle of important datasets easier to understand and document.<\/span><\/p>\nQuestion 396. What is an idempotent operation in an ML data pipeline?<\/b><\/h3>\n
1)<\/b> An operation that always produces random output<\/span>
\n<\/span>2)<\/b> An operation that requires manual execution every time<\/span>
\n<\/span>3)<\/b> An operation that changes its result whenever it is repeated<\/span>
\n<\/span>4)<\/b> An operation that can be repeated without causing unintended additional effects**<\/span><\/p>\nAnswer: 4) An operation that can be repeated without causing unintended additional effects<\/b><\/p>\n
Explanation:<\/b><\/p>\n
An idempotent operation produces the same intended final state when it is executed repeatedly with the same relevant inputs. This property is valuable in machine learning pipelines because workflows may be retried after failures, network interruptions, or temporary service problems. For example, a processing step that safely overwrites a known output rather than creating multiple unintended copies can be easier to retry. Idempotency can improve pipeline reliability and simplify recovery procedures. Not every operation is naturally idempotent, so workflow designers may need to use unique identifiers, controlled writes, or state checks to prevent duplicate or unintended side effects during retries.<\/span><\/p>\nQuestion 397. What is a dead-letter queue commonly used for in an event-driven ML workflow?<\/b><\/h3>\n
1)<\/b> To permanently increase model accuracy<\/span>
\n<\/span>2)<\/b> To store messages that could not be successfully processed<\/span>
\n<\/span>3)<\/b> To replace the training dataset<\/span>
\n<\/span>4)<\/b> To compress model parameters<\/span><\/p>\nAnswer: 2) To store messages that repeatedly fail processing<\/b><\/p>\n
Explanation:<\/b><\/p>\n
A dead-letter queue, or DLQ, can store messages that cannot be successfully processed after a configured number of attempts. In an event-driven machine learning workflow, this can help prevent problematic messages from repeatedly blocking normal processing. For example, a malformed inference request or invalid event may fail processing multiple times and eventually be moved to the DLQ for investigation. The failed message can then be inspected, corrected, replayed, or otherwise handled according to the application’s operational process. A DLQ does not itself correct the underlying problem; it provides a controlled location for handling messages that require additional attention.<\/span><\/p>\nQuestion 398. What is the purpose of retry logic in an ML pipeline?<\/b><\/h3>\n
1)<\/b> To recover automatically from appropriate transient failures<\/span>
\n<\/span>2)<\/b> To guarantee every model prediction is correct<\/span>
\n<\/span>3)<\/b> To remove all validation checks<\/span>
\n<\/span>4)<\/b> To permanently ignore failed operations<\/span><\/p>\n