View Full Amazon AWS Certified AI Practitioner AIF-C01 Exam Dumps and Practice Test Dumps.

 

Question 161

A machine learning model performs poorly on both its training data and unseen data because it is too simple to capture important patterns. What problem does this describe?

1. Underfitting
2. Data leakage
3. Overfitting
4. Data poisoning

Correct Answer: 1

Explanation

Underfitting occurs when a machine learning model is too simple to adequately learn the important patterns in its training data. As a result, the model can perform poorly on both training and unseen data. Underfitting may result from an overly simple model, insufficient training, overly restrictive assumptions, or inadequate features. Increasing model complexity, improving feature representation, providing better training data, or adjusting training procedures may help address the problem. This differs from overfitting, where a model learns training data too closely and performs poorly when presented with new data.

Question 162

Which technique divides available training data into multiple subsets so that a model can be trained and evaluated repeatedly using different portions of the data?

1. Tokenization
2. Cross-validation
3. Quantization
4. Embedding

Correct Answer: 2

Explanation

Cross-validation is a model evaluation technique that divides a dataset into multiple subsets, commonly called folds. The model is trained on some folds and evaluated on another, with the process repeated so different portions of the data serve as validation data. The resulting measurements can provide a more robust estimate of model performance than relying on a single split. Cross-validation is especially useful when available data is limited. Care must be taken to prevent information leakage between training and validation portions, and a separate test dataset may still be retained for final evaluation.

Question 163

During neural network training, which parameter controls the size of the updates made to model parameters during optimization?

1. Batch size
2. Epoch count
3. Learning rate
4. Context window

Correct Answer: 3

Explanation

The learning rate controls the magnitude of parameter updates during model optimization. A learning rate that is too large can cause training to become unstable or overshoot useful solutions, while a rate that is too small can make training unnecessarily slow. The appropriate value depends on the model, optimizer, dataset, and training configuration. Learning-rate schedules can also adjust the rate during training. The learning rate is different from batch size, which determines how many training examples are processed in one update. Proper training configuration typically requires experimentation and validation.

Question 164

A data scientist wants to transform numerical features so that differences in their original measurement scales have less influence on a model. Which preprocessing approach is appropriate?

1. Data labeling
2. Feature extraction
3. Model pruning
4. Feature scaling

Correct Answer: 4

Explanation

Feature scaling transforms numerical variables so their magnitudes are placed on comparable scales. This can be particularly useful for algorithms that are sensitive to feature magnitude or rely on distance calculations. Common approaches include normalization and standardization. For example, a feature measured in thousands could otherwise have a much larger numerical influence than another feature measured between zero and one. Scaling parameters should generally be calculated from training data and then consistently applied to validation, test, and production data. Proper scaling can support more stable and effective model training.

Question 165

An ML dataset contains missing values in a numerical field. Which action can be used to prepare the data before training, depending on the application’s requirements?

1. Replace or otherwise handle missing values using an appropriate imputation strategy
2. Automatically delete the entire machine learning system
3. Convert every numerical value into an image
4. Increase the model’s output temperature

Correct Answer: 1

Explanation

Missing numerical values can be handled using an appropriate imputation strategy or another method suitable for the dataset and application. Common approaches may include replacing missing values with statistical estimates such as a mean or median, although the best method depends on the data distribution and business meaning. In some cases, records with missing values may be removed or a model may support missing values directly. The chosen approach should be determined using training data and applied consistently. Teams should also investigate why values are missing because missingness itself can contain useful information.

Question 166

What is the primary purpose of gradient descent in machine learning training?

1. To encrypt model parameters
2. To adjust model parameters in a direction that reduces the training loss
3. To convert text into speech
4. To increase the number of input tokens

Correct Answer: 2

Explanation

Gradient descent is an optimization method used to adjust model parameters with the goal of reducing a selected loss function. The algorithm uses information about the gradient of the loss to determine the direction in which parameters should be changed. Repeated updates can move the model toward parameter values that produce lower loss on the training objective. Variants such as stochastic and mini-batch gradient descent are commonly used in machine learning. The learning rate influences the size of these updates and therefore plays an important role in training behavior.

Question 167

An organization wants to maintain approved versions of machine learning models and track which version is ready for deployment. Which capability is most directly suited to this requirement?

1. Amazon Route 53
2. Amazon S3 Glacier
3. Amazon SageMaker Model Registry
4. AWS Direct Connect

Correct Answer: 3

Explanation

Amazon SageMaker Model Registry is designed to help organizations manage machine learning model versions and organize models through a controlled lifecycle. Teams can register models, associate metadata and evaluation information, and manage which versions are approved for deployment. This supports governance and reproducibility by providing a structured way to track model artifacts and lifecycle status. Model management is especially useful when multiple versions are tested or deployed across environments. Organizations can combine model registry capabilities with evaluation, monitoring, access controls, and deployment workflows to establish a more controlled machine learning lifecycle.

Question 168

A company uses a generative AI application to retrieve information from several sources. It wants to improve results by combining keyword matching with semantic similarity. Which search approach can support this requirement?

1. Hybrid search
2. Random search
3. Image segmentation
4. Regression analysis

Correct Answer: 1

Explanation

Hybrid search combines different retrieval approaches, such as traditional keyword-based search and semantic or vector-based search. Keyword matching can perform well when users provide specific terms, names, or identifiers, while semantic search can help retrieve content based on meaning even when exact wording differs. Combining these methods can improve retrieval quality for applications that need to find relevant information from large document collections. The appropriate balance depends on the data and workload. Retrieval results should still be evaluated using representative queries to determine whether the approach provides sufficiently relevant context.

Question 169

In a retrieval-augmented generation application, what is the purpose of a reranking step after an initial retrieval operation?

1. To encrypt all retrieved documents
2. To replace the foundation model
3. To reorder retrieved results so the most relevant items appear first
4. To remove user authentication

Correct Answer: 3

Explanation

Reranking is used to improve the ordering of results returned by an initial retrieval stage. A retrieval system may first identify a broader set of potentially relevant documents efficiently. A reranker can then evaluate those candidates in greater detail and place the most relevant items near the top. This can help a generative AI model receive higher-quality context when only a limited amount of retrieved information can be included. Reranking does not replace authentication or encryption, and it does not itself generate the final response. Its primary role is improving retrieval relevance.

Question 170

Which AWS capability is specifically designed to help applications use foundation models together with organizational data stored in knowledge sources?

1. Amazon Bedrock Knowledge Bases
2. Amazon EC2 Auto Scaling
3. Amazon Route 53 Resolver
4. AWS Elastic Beanstalk

Correct Answer: 1

Explanation

Amazon Bedrock Knowledge Bases provides capabilities for connecting generative AI applications with organizational information so relevant content can be retrieved and supplied to a foundation model. This supports retrieval-augmented generation patterns in which responses can be grounded using information from configured data sources. Knowledge retrieval can help applications use current organizational content without requiring all information to be encoded directly into a model’s parameters. Organizations should still manage permissions, data quality, synchronization, retrieval relevance, and evaluation to ensure that generated responses are based on appropriate source information.

Question 171

A development team wants to make a smaller model perform a task using knowledge or behavior learned from a larger model. Which technique can support this goal?

1. Data masking
2. Knowledge distillation
3. Data deletion
4. Network segmentation

Correct Answer: 2

Explanation

Knowledge distillation is a technique in which a smaller student model learns behavior from a larger teacher model. The goal can be to produce a model that requires fewer computational resources while retaining useful capabilities of the larger model. The student may learn from outputs or other information produced by the teacher. Distillation can be useful when deployment constraints require lower latency, reduced resource consumption, or a smaller model footprint. The resulting model should still be evaluated against the intended workload because reducing model size can affect performance and capability.

Question 172

Which concern is particularly important when an organization uses personally identifiable information as input to an AI application?

1. Choosing a larger monitor
2. Increasing screen brightness
3. Protecting the information through appropriate privacy and access controls
4. Changing the application’s font

Correct Answer: 3

Explanation

Personally identifiable information, or PII, requires appropriate protection because its disclosure or misuse can create privacy and compliance risks. Organizations should determine whether collecting the information is necessary, restrict access to authorized users and systems, and apply suitable security controls. Depending on the use case, techniques such as masking, redaction, anonymization, or pseudonymization may also be appropriate. Data should be handled according to applicable policies and legal requirements. AI applications require the same careful data governance principles as other systems, with additional consideration for how prompts, retrieved context, outputs, and logs are handled.

Question 173

What is the main purpose of model checkpointing during a machine learning training process?

1. To save intermediate model states so training can be resumed or selected states can be retained
2. To convert numerical data into text
3. To assign IAM permissions
4. To increase network bandwidth

Correct Answer: 1

Explanation

Model checkpointing saves the state of a model during training at selected points. A checkpoint can contain model parameters and, depending on the training configuration, additional information needed to continue training. Checkpoints can help recover from interruptions and allow teams to retain useful intermediate model states. They can also support experimentation by enabling developers to evaluate different points in the training process. Checkpoint management should consider storage requirements, versioning, security, and retention policies. Checkpointing does not itself guarantee that a saved model performs well; saved states still require appropriate evaluation.

Question 174

An organization wants to determine whether an AI model has been trained using data that may contain unauthorized or inappropriate content. Which governance activity is most relevant?

1. Increasing inference temperature
2. Reviewing data provenance and lineage
3. Increasing the model’s context window
4. Disabling audit logs

Correct Answer: 2

Explanation

Data provenance and lineage help organizations understand where data originated, how it was obtained, and how it moved or was transformed throughout a data workflow. Reviewing provenance can support governance decisions about whether data was collected and used appropriately. It can also help organizations investigate problems involving data quality, licensing, privacy, or unauthorized sources. For AI workloads, understanding training and retrieval data sources can be important when assessing compliance and responsible use. Provenance information should be maintained through appropriate processes and documentation so organizations can trace important datasets and transformations.

Question 175

Which statement best describes reinforcement learning from human feedback (RLHF)?

1. A method that uses human preference information to help optimize model behavior
2. A technique used only to compress image files
3. A database indexing method
4. A network routing protocol

Correct Answer: 1

Explanation

Reinforcement learning from human feedback, commonly called RLHF, uses human preference information to help guide a model toward desired behavior. Human reviewers may compare or rank candidate responses, and this information can be incorporated into a training or optimization process. RLHF is one approach used in developing and aligning certain generative AI systems. It does not mean that humans manually write every model response. Human feedback provides signals about preferred behavior, while the underlying training process uses those signals to optimize the model. The specific implementation can vary between model-development workflows.

Question 176

A company wants to use a model for an application that requires both text and image inputs. Which model-selection factor should receive particular attention?

1. The model’s supported input modalities
2. The color of the application dashboard
3. The number of office printers
4. The database table naming convention

Correct Answer: 1

Explanation

Model selection should consider whether the model supports the input and output modalities required by the application. A workload requiring both text and image inputs needs a model with appropriate multimodal capabilities. Selecting a model solely based on general performance or size may result in an incompatible solution. Organizations should also evaluate factors such as quality, latency, cost, safety, context requirements, and supported features. Model capabilities can differ significantly, so teams should verify the documentation for the specific model version and test it using representative application inputs before deployment.

Question 177

Which practice can help an organization maintain traceability when multiple versions of an AI model are deployed over time?

1. Recording model versions and associated deployment information
2. Removing all deployment records
3. Allowing unidentified model artifacts
4. Reusing the same version identifier for every release

Correct Answer: 1

Explanation

Recording model versions and associated deployment information supports traceability across the AI lifecycle. Organizations can use version identifiers, artifact records, deployment timestamps, evaluation results, and related metadata to determine which model was used at a particular time. This can help with troubleshooting, audits, reproducibility, and controlled rollback. Reusing identifiers or deleting deployment records makes it more difficult to investigate historical behavior. Model versioning is particularly important when models are updated frequently or when AI outputs influence significant business processes. Traceability should be integrated into the organization’s broader governance practices.

Question 178

A generative AI application is designed to perform actions in external systems. Which design principle is most appropriate before allowing a model-generated action to execute?

1. Allow every action automatically
2. Validate the requested action and enforce authorization before execution
3. Remove all identity controls
4. Give the model unrestricted administrative privileges

Correct Answer: 2

Explanation

Model-generated actions should be validated and authorized before they are executed against external systems. A generative model can produce incorrect, unexpected, or manipulated instructions, so applications should not treat generated content as inherently trustworthy commands. The application should verify that the requested operation is valid, confirm that the user or service is authorized, and apply appropriate safeguards. High-impact actions may also require additional approval or human review. This design separates generation from execution and places security controls in the application and infrastructure layers rather than relying solely on model behavior.

Question 179

Which AWS service can help organizations analyze and identify sensitive data such as personally identifiable information stored in Amazon S3?

1. Amazon Macie
2. Amazon Polly
3. Amazon Lex
4. Amazon Translate

Correct Answer: 1

Explanation

Amazon Macie is an AWS service designed to help discover and protect sensitive data stored in Amazon S3. It can identify and provide findings related to sensitive information, including certain types of personally identifiable information, helping organizations understand where sensitive data exists. This visibility can support data governance and security programs for AI workloads that use S3 data. Organizations can use the findings to investigate storage configurations, access patterns, and data-handling practices. Macie does not replace access controls or encryption; it complements broader security and data-protection measures.

Question 180

A company wants an AI assistant to answer employees’ business questions while respecting the access permissions associated with corporate information. What capability is particularly important?

1. Permission-aware retrieval
2. Unrestricted document access
3. Anonymous data retrieval
4. Global administrator privileges

Correct Answer: 1

Explanation

Permission-aware retrieval ensures that an AI assistant retrieves information only from sources the requesting user is authorized to access. This is especially important for enterprise assistants because corporate repositories can contain confidential information belonging to different teams or roles. Without appropriate authorization filtering, a model could potentially expose information through an otherwise valid response. Access controls should therefore be enforced by the application or retrieval system rather than relying solely on the model to decide what information is appropriate. Organizations should test permission boundaries carefully to verify that unauthorized content cannot enter generated responses.