AWS ML Specialty（English scenarios）

Batch Transform vs endpoint, concept drift, MSK vs Kinesis, tensor parallelism. Longer English stems, not a gloss of the Japanese set.

同梱問題集の公開抜粋です。最初は問題と選択肢だけ出します。正解と解説は各問の「正解と解説を見る」を開くまで見えません。作者の既定表示は Ultra Practicebook、ライセンスは MIT です。試験団体の過去問や市販問題集の転載ではありません。

この問題集の位置づけ

This is not a translation of the Japanese AWS set. The Japanese questions pick a service or instance family from a short definition. These English items keep a constraint in the stem: you do not want a 24/7 endpoint, you already run Kafka, you must catch a change in p(y|x), a single layer does not fit on one GPU.

Random Cut Forest is here because it is the built-in detector for tabular anomalies; XGBoost in the same list is a supervised tree booster and does not fill that role. Autopilot's default split is a fact to look up once, then stop memorizing console trivia that changes.

These items are original. They follow public AWS documentation, not live exam dumps.

学習目標

Choose a service from constraints in a short scenario: cost of a persistent endpoint, Kafka compatibility, column-level S3 access, detecting a change in the label relationship rather than the input distribution.

対象者

Readers who already know service names in Japanese or from the console, and now need English exam-style stems. People who confuse data drift with concept drift, or data parallelism with tensor parallelism.

学習方法

Underline the constraint in the stem (cost, latency, Kafka API, GPU memory) before looking at options. After an error, write one sentence that names the rejected service and the constraint it fails. Keep the Japanese set for name-to-purpose drills in a different session.

問題

先に自分で選んでから、折りたたみを開いてください。解説は正解の言い換えだけでなく、隣の選択肢が文に入らない理由や、一緒に覚える連語を書いています。番号だけ暗記せず、解説の文を一度声に出してから次へ進んでください。

問 1

Which SageMaker built-in algorithm is specifically designed for anomaly detection on tabular data using an unsupervised approach?

  1. XGBoost
  2. Random Cut Forest
  3. BlazingText
  4. DeepAR+
正解と解説を見る

正解: Random Cut Forest

Random Cut Forest (RCF) is SageMaker's built-in unsupervised algorithm for anomaly detection. It assigns an anomaly score to each data point based on how much adding it to a random cut tree changes the tree complexity. The stem asks: Which SageMaker built-in algorithm is specifically designed for anomaly detection on tabul. The keyed answer is 2. Options: XGBoost; Random Cut Forest; BlazingText; DeepAR+. Keep this in the AWS bucket, and reject the neighbor term that solves a different job.

問 2

In Amazon SageMaker, what is the purpose of a Model Package in the Model Registry?

  1. A Docker image for training containers
  2. A versioned deployable artifact that bundles model artifacts and inference code
  3. A managed feature group stored in Feature Store
  4. A pipeline step definition for data preprocessing
正解と解説を見る

正解: A versioned deployable artifact that bundles model artifacts and inference code

SageMaker Model Registry organizes trained models into Model Packages with version control. Each package stores the S3 path to model artifacts and container image URI, enabling approval workflows before production deployment.

問 3

Which AWS service provides a fully managed Apache Kafka-compatible streaming platform?

  1. Amazon Kinesis Data Streams
  2. Amazon Kinesis Data Firehose
  3. Amazon MSK (Managed Streaming for Apache Kafka)
  4. AWS Glue Streaming
正解と解説を見る

正解: Amazon MSK (Managed Streaming for Apache Kafka)

Amazon MSK is a fully managed service that runs Apache Kafka. Unlike Kinesis which is proprietary, MSK is Kafka-compatible and suits teams with existing Kafka expertise or tooling. The stem asks: Which AWS service provides a fully managed Apache Kafka-compatible streaming platform?. The keyed answer is 3. Options: Amazon Kinesis Data Streams; Amazon Kinesis Data Firehose; Amazon MSK (Managed Streaming for Apache Kafka); AWS Glue Streaming. Keep this in the AWS bucket, and reject the neighbor term that solves a different job.

問 4

What is the default split ratio used by SageMaker Autopilot when dividing data into training and validation sets?

  1. 90% training / 10% validation
  2. 80% training / 20% validation
  3. 70% training / 30% validation
  4. 60% training / 40% validation
正解と解説を見る

正解: 80% training / 20% validation

SageMaker Autopilot defaults to an 80/20 train-validation split unless the user specifies a custom ratio. This is a widely used convention that balances model training data volume with reliable evaluation. The stem asks: What is the default split ratio used by SageMaker Autopilot when dividing data into traini. The keyed answer is 2. Options: 90% training / 10% validation; 80% training / 20% validation; 70% training / 30% validation; 60% training / 40% validation. Keep this in the AWS bucket, and reject the neighbor term that solves a different job.

問 5

When using SageMaker Batch Transform for inference on a large dataset stored in S3, what is the key advantage over using a real-time endpoint?

  1. Lower per-record latency
  2. No need to provision and maintain a persistent endpoint
  3. Support for streaming input
  4. Built-in A/B testing
正解と解説を見る

正解: No need to provision and maintain a persistent endpoint

Batch Transform spins up compute only for the duration of the job and terminates it afterward, making it cost-efficient for offline bulk inference. Real-time endpoints remain running and incur continuous costs. The stem asks: When using SageMaker Batch Transform for inference on a large dataset stored in S3, what i. The keyed answer is 2. Options: Lower per-record latency; No need to provision and maintain a persistent endpoint; Support for streaming input; Built-in A/B testing. Keep this in the AWS bucket, and reject the neighbor term that solves a different job.

問 6

In the context of AWS ML, what does the term 'concept drift' refer to?

  1. A change in model architecture during retraining
  2. A shift in the statistical relationship between input features and the target variable over time
  3. An increase in feature dimensionality after feature engineering
  4. A mismatch between training data schema and inference data schema
正解と解説を見る

正解: A shift in the statistical relationship between input features and the target variable over time

Concept drift occurs when the underlying relationship the model learned changes over time (e.g., user behavior shifts). It differs from data drift which is purely a change in input distribution. SageMaker Model Monitor can detect both.

問 7

Which SageMaker feature allows you to interactively prepare and visualize data without writing code?

  1. SageMaker Studio Notebooks
  2. SageMaker Data Wrangler
  3. SageMaker Processing Jobs
  4. AWS Glue DataBrew
正解と解説を見る

正解: SageMaker Data Wrangler

SageMaker Data Wrangler provides a GUI for data import, visualization, transformation, and analysis with 300+ built-in transforms. It can export the transformation flow to a SageMaker Pipeline step. The stem asks: Which SageMaker feature allows you to interactively prepare and visualize data without wri. The keyed answer is 2. Options: SageMaker Studio Notebooks; SageMaker Data Wrangler; SageMaker Processing Jobs; AWS Glue DataBrew. Keep this in the AWS bucket, and reject the neighbor term that solves a different job.

問 8

What type of parallelism does SageMaker's Tensor Parallelism strategy address?

  1. Distributing different mini-batches across GPUs
  2. Splitting individual tensor operations (e.g. weight matrices) across multiple GPUs
  3. Running separate hyperparameter tuning trials in parallel
  4. Processing multiple inference requests simultaneously
正解と解説を見る

正解: Splitting individual tensor operations (e.g. weight matrices) across multiple GPUs

Tensor parallelism partitions large weight matrices across GPUs so that each GPU holds only a shard. This is critical for training LLMs whose single layers exceed the memory of one GPU. It differs from data parallelism which replicates the full model.

問 9

Amazon Personalize uses which type of machine learning technique as its core approach?

  1. Supervised classification with gradient boosting
  2. Collaborative filtering and neural network-based ranking
  3. Rule-based expert systems
  4. Generative adversarial networks
正解と解説を見る

正解: Collaborative filtering and neural network-based ranking

Amazon Personalize is built on algorithms including collaborative filtering and deep learning ranking models (similar to those used at Amazon.com). It ingests user-item interaction data to generate personalized recommendations.

問 10

In SageMaker, which container environment variable specifies the path where the training script should write model artifacts?

  1. SM_MODEL_DIR
  2. SM_CHANNEL_TRAINING
  3. SM_OUTPUT_DATA_DIR
  4. SM_HP_EPOCHS
正解と解説を見る

正解: SM_MODEL_DIR

SM_MODEL_DIR (default: /opt/ml/model) is where the training script saves the final model artifacts. SageMaker tars and uploads this directory to S3 after training completes. SM_CHANNEL_TRAINING points to input data. The stem asks: In SageMaker, which container environment variable specifies the path where the training s. The keyed answer is 1. Options: SM_MODEL_DIR; SM_CHANNEL_TRAINING; SM_OUTPUT_DATA_DIR; SM_HP_EPOCHS. Keep this in the AWS bucket, and reject the neighbor term that solves a different job.

問 11

The F1 score is defined as the harmonic mean of precision and recall.

正解と解説を見る

正解: true

F1 = 2 * (Precision * Recall) / (Precision + Recall). Using the harmonic mean penalizes extreme imbalance between precision and recall more than the arithmetic mean would. The stem asks: The F1 score is defined as the harmonic mean of precision and recall.. The keyed answer is true. Keep this in the ML bucket, and reject the neighbor term that solves a different job.

問 12

Amazon SageMaker supports deploying models trained outside of AWS (e.g. locally trained PyTorch models) to SageMaker Endpoints.

正解と解説を見る

正解: true

SageMaker allows you to bring your own model by uploading model artifacts to S3 and specifying a compatible inference container (AWS pre-built or custom). This is the 'Bring Your Own Model' (BYOM) pattern. The stem asks: Amazon SageMaker supports deploying models trained outside of AWS (e.g. locally trained Py. The keyed answer is true. Keep this in the AWS bucket, and reject the neighbor term that solves a different job.

問 13

AWS Lake Formation can be used to define fine-grained column-level access controls on data stored in Amazon S3.

正解と解説を見る

正解: true

Lake Formation supports table-, column-, row-, and cell-level security on S3-backed data catalogs. Permissions are enforced when data is accessed through AWS Glue, Athena, or Amazon Redshift Spectrum. The stem asks: AWS Lake Formation can be used to define fine-grained column-level access controls on data. The keyed answer is true. Keep this in the AWS bucket, and reject the neighbor term that solves a different job.

問 14

SageMaker Processing Jobs can only run Python scripts using the built-in Scikit-learn container.

正解と解説を見る

正解: false

SageMaker Processing Jobs support multiple frameworks via pre-built containers (Scikit-learn, Spark, XGBoost) and also support custom Docker containers, allowing arbitrary code in any language. The stem asks: SageMaker Processing Jobs can only run Python scripts using the built-in Scikit-learn cont. The keyed answer is false. Keep this in the AWS bucket, and reject the neighbor term that solves a different job.

問 15

In transformer-based LLMs, increasing the number of attention heads while keeping total model dimensions fixed will increase per-head dimensionality.

正解と解説を見る

正解: false

With fixed model dimension d_model, head dimension = d_model / num_heads. Increasing num_heads while d_model is fixed reduces per-head dimensionality, not increases it. The stem asks: In transformer-based LLMs, increasing the number of attention heads while keeping total mo. The keyed answer is false. Keep this in the ML bucket, and reject the neighbor term that solves a different job.

問 16

What is the name of the AWS service that provides a vector database capability for similarity search used in RAG architectures?

正解と解説を見る

正解: Amazon OpenSearch Service

Amazon OpenSearch Service supports k-NN (k-nearest neighbor) vector search using FAISS and NMSLIB engines, making it a common choice for storing and querying embeddings in Retrieval-Augmented Generation pipelines. The stem asks: What is the name of the AWS service that provides a vector database capability for similar. The keyed answer is Amazon OpenSearch Service. Keep this in the AWS bucket, and reject the neighbor term that solves a different job.

問 17

What algorithm does Amazon Fraud Detector use as its primary modeling approach?

正解と解説を見る

正解: Amazon Fraud Detector uses a combination of supervised ML models including logistic regression and neural networks trained on labeled fraud and legitimate transaction data.

Amazon Fraud Detector allows users to bring labeled fraud data and automatically trains supervised models. It supports rule-based logic in combination with ML predictions for flexible fraud decisioning. The stem asks: What algorithm does Amazon Fraud Detector use as its primary modeling approach?. The keyed answer is Amazon Fraud Detector uses a combination of supervised ML models including logistic regression and neural networks trained on labeled fraud and legitimate transaction data.. Keep this in the AWS bucket, and reject the neighbor term that solves a different job.

問 18

Which AWS compute instance family is equipped with AWS Trainium chips designed specifically for deep learning training?

  1. ml.p4 instances
  2. ml.trn1 instances
  3. ml.inf1 instances
  4. ml.g5 instances
正解と解説を見る

正解: ml.trn1 instances

ml.trn1 instances are powered by AWS Trainium, a purpose-built ML training chip. ml.inf1 uses Inferentia for inference, ml.p4 uses NVIDIA A100 GPUs, and ml.g5 uses NVIDIA A10G GPUs. The stem asks: Which AWS compute instance family is equipped with AWS Trainium chips designed specificall. The keyed answer is 2. Options: ml.p4 instances; ml.trn1 instances; ml.inf1 instances; ml.g5 instances. Keep this in the AWS bucket, and reject the neighbor term that solves a different job.

問 19

In the CRISP-DM methodology, which phase directly precedes modeling?

  1. Business Understanding
  2. Data Understanding
  3. Data Preparation
  4. Evaluation
正解と解説を見る

正解: Data Preparation

CRISP-DM phases in order: Business Understanding -> Data Understanding -> Data Preparation -> Modeling -> Evaluation -> Deployment. Data Preparation (cleaning, transforming, feature engineering) must be completed before model training.

問 20

Which metric is most appropriate for evaluating a recommendation system that returns a ranked list of items?

  1. Accuracy
  2. Mean Squared Error
  3. Normalized Discounted Cumulative Gain (NDCG)
  4. Area Under the ROC Curve
正解と解説を見る

正解: Normalized Discounted Cumulative Gain (NDCG)

NDCG measures the quality of a ranked list by rewarding relevant items appearing higher in the list (applying a logarithmic discount). It is a standard metric for ranking and recommendation tasks. The stem asks: Which metric is most appropriate for evaluating a recommendation system that returns a ran. The keyed answer is 3. Options: Accuracy; Mean Squared Error; Normalized Discounted Cumulative Gain (NDCG); Area Under the ROC Curve. Keep this in the ML bucket, and reject the neighbor term that solves a different job.