MLS-C01 Practice Exam 56 — 20 Questions

Question 1

A team needs to automatically retrain a model every week using new data. Which SageMaker feature is designed to schedule and automate this workflow?

Accepted Answer

SageMaker Pipelines. SageMaker Pipelines is a purpose-built CI/CD service for machine learning that lets you define, schedule, and automate multi-step workflows including data preprocessing, training, evaluation, and model registration. You can trigger a pipeline on a weekly schedule using EventBridge or the Pipelines scheduling capability, making it the correct choice for automated weekly retraining.

Answer

SageMaker Automatic Model Tuning

Answer

SageMaker Model Monitor

Answer

SageMaker Data Wrangler

Question 2

A data scientist is evaluating a binary classification model. The model's confusion matrix shows: True Positives=80, False Positives=20, True Negatives=900, False Negatives=0. Which THREE metrics can be calculated from this confusion matrix? (Choose three.)

Accepted Answer

Precision. Precision (Option A) is correct because it is computed directly from the confusion matrix as TP/(TP+FP) = 80/(80+20) = 0.80, using only the true positive and false positive counts. Recall (Option B) is correct because it equals TP/(TP+FN) = 80/(80+0) = 1.00, requiring only true positives and false negatives from the matrix. Accuracy (Option D) is correct because it is (TP+TN)/(TP+TN+FP+FN) = (80+900)/1000 = 0.98, which uses all four cells of the confusion matrix. AUC-ROC (Option C) is not calculable here because it requires the model's continuous predicted scores or probabilities across many decision thresholds, not just a single threshold's confusion matrix counts. RMSE (Option E) does not belong because it is a regression error metric computed from continuous predicted-versus-actual numeric residuals, whereas this is a binary classification scenario with only categorical counts.

Answer

AUC-ROC

Answer

Root Mean Squared Error (RMSE)

Question 3

A team is training a neural network for image classification using Amazon SageMaker. The training loss decreases rapidly but the validation loss starts increasing after a few epochs. Which action should the team take?

Accepted Answer

Implement early stopping based on validation loss. The described behavior — training loss decreasing while validation loss rises — is the textbook signature of overfitting. Early stopping monitors validation loss during training and halts the job once validation performance stops improving (or begins to degrade), preserving the best model checkpoint. This directly addresses the observed symptom without altering the model architecture or optimization dynamics.

Answer

Reduce the batch size

Answer

Add more convolutional layers

Answer

Increase the learning rate

Question 4

A company is using SageMaker to train a deep learning model with TensorFlow. The training job is running on an ml.p3.16xlarge instance. The data scientist wants to maximize GPU utilization. Which configuration should be used?

Accepted Answer

Use Pipe mode or Fast File mode with a large batch size that fits in GPU memory.. Pipe mode and Fast File mode stream training data directly from Amazon S3 into the training container without downloading the full dataset to the instance's EBS volume first, which removes the I/O bottleneck that starves GPUs on large datasets. Combining this with the largest batch size that fits in GPU memory keeps the GPU compute units saturated with matrix operations instead of waiting on data. On an ml.p3.16xlarge with 8 NVIDIA V100 GPUs, this is the standard SageMaker pattern for maximizing GPU utilization.

Answer

Use a single GPU and increase the number of epochs.

Answer

Use a CPU-only instance for training and then deploy on GPU.

Answer

Use File mode input and a small batch size.

Question 5

A machine learning team is using Amazon SageMaker to train a PyTorch model on a dataset that is 500 GB in size. The training job runs on a single ml.p3.2xlarge instance, but the training takes over 48 hours, which exceeds the maximum allowed time. The team wants to reduce training time to under 24 hours. They are open to using multiple instances and have budget for up to 4 instances. The dataset is stored in Amazon S3 and can be split into shards by a key. The model architecture must remain unchanged. What should the team do?

Accepted Answer

Use SageMaker distributed data parallelism with 4 ml.p3.2xlarge instances.. SageMaker distributed data parallelism shards the dataset across multiple instances and uses an all-reduce communication pattern to synchronize gradients, allowing the same model architecture to train on 4 instances in parallel. With the dataset already shardable by key and budget for 4 instances, this is the correct approach to cut training time while keeping the model unchanged. It scales throughput roughly linearly with the number of instances for data-parallel workloads.

Answer

Use SageMaker Processing to split the data and train separate models.

Answer

Change the instance type to ml.p3.16xlarge.

Answer

Switch to Pipe input mode to stream data faster.

Question 6

A team is using Amazon SageMaker to train a deep learning model for image classification. The training job is taking too long, and they want to reduce training time without sacrificing model accuracy. Which approach is most effective?

Accepted Answer

Use transfer learning with a pre-trained model and fine-tune on the target dataset. Transfer learning with a pre-trained model leverages weights already learned on a large dataset (e.g., ImageNet) and fine-tunes them on the target dataset, dramatically reducing the number of epochs and compute needed to reach high accuracy. This is the most effective approach because the model starts from a strong feature representation rather than random initialization. It preserves or improves accuracy while cutting training time.

Answer

Reduce the batch size

Answer

Reduce the number of training epochs

Answer

Reduce the image resolution

Question 7

Which THREE techniques can help reduce overfitting in a neural network? (Select THREE.)

Accepted Answer

Dropout. Dropout (B) is correct because randomly deactivating a fraction of neurons during each training iteration prevents the network from relying on specific units and forces redundant, more generalizable feature representations. Early stopping (C) is correct because it halts training when validation loss stops improving, preventing the model from continuing to fit noise in the training data. L2 regularization (E) is correct because adding a penalty proportional to the squared magnitude of weights (weight decay) constrains the model's parameters and reduces variance. Increasing training epochs (A) is not correct because more epochs typically allow the network to fit the training data even more closely, worsening overfitting. Increasing the number of layers (D) is not correct because adding depth increases model capacity, which makes overfitting more likely rather than less.

Answer

Increase training epochs

Answer

Increase the number of layers

Question 8

A machine learning team stores 40 TB of Parquet training data in Amazon S3 and wants to query and transform it with Amazon Athena before training. Queries currently scan the entire dataset and cost is rising. The data is partitioned by year, month, and day in the S3 prefix structure, but query performance has not improved. A data engineer confirms the table was created with AWS Glue Data Catalog, but the partition metadata is missing. What should the engineer do to enable partition pruning and reduce scanned data?

Accepted Answer

Run MSCK REPAIR TABLE on the Athena table to load the partition metadata into the AWS Glue Data Catalog.. Registering the existing Hive-style partitions in the Data Catalog is the key. MSCK REPAIR TABLE detects the year/month/day prefix structure and adds partition entries, allowing Athena to prune non-matching partitions and scan only relevant data. Changing formats or enabling transfer features does not address the missing metadata that prevents pruning.

Answer

Convert the Parquet files to CSV and re-crawl them with an AWS Glue crawler to force partition discovery.

Answer

Enable Amazon S3 Transfer Acceleration on the bucket to speed up Athena reads.

Answer

Add a WHERE clause filtering on the S3 object key prefix instead of on partition columns.

Question 9

A machine learning team needs to preprocess large volumes of clickstream data stored in Amazon S3 before training a model. The preprocessing includes data cleaning, feature engineering, and normalization. The team wants to use a serverless solution that minimizes operational overhead. Which combination of services should the team use?

Accepted Answer

AWS Glue ETL jobs reading from and writing to S3.. AWS Glue ETL jobs are a serverless solution that automatically provisions and scales the underlying compute resources, making them ideal for preprocessing large volumes of clickstream data stored in S3. Glue can read directly from S3, perform data cleaning, feature engineering, and normalization using PySpark or Scala, and write the transformed data back to S3, all without managing any infrastructure. This minimizes operational overhead while handling the required preprocessing tasks at scale.

Answer

Amazon SageMaker Notebooks with custom Python scripts.

Answer

Amazon EMR with Spark clusters.

Answer

Amazon Athena with SQL queries.

Question 10

Which TWO actions can help reduce overfitting in a decision tree model? (Choose 2.)

Accepted Answer

Prune the tree after training. Option A, pruning the tree after training, is correct because post-pruning (e.g., cost-complexity pruning with ccp_alpha in scikit-learn) removes branches that add little predictive power, reducing variance and thus overfitting. Option C, setting a minimum number of samples per leaf (min_samples_leaf), is correct because it prevents the tree from creating leaves that fit noise from very few samples, forcing more generalizable splits. Option B is wrong because increasing max_depth lets the tree grow deeper and fit training noise more closely, worsening overfitting. Option D is wrong because increasing the number of features considered at each split (max_features) tends to make splits more specific to training data, not less. Option E is wrong because training on all data without a validation set removes the ability to detect or tune against overfitting.

Answer

Increase the maximum depth of the tree

Answer

Increase the number of features considered at each split

Answer

Use all training data without validation

Question 11

A machine learning team is using Amazon SageMaker to train a model with a large dataset stored in S3. The training job is taking too long. Which THREE of the following actions can reduce training time? (Choose three.)

Accepted Answer

Use a GPU instance with more powerful GPUs.. Option B is correct because selecting a SageMaker training instance type with more powerful GPUs (e.g., moving from ml.p3.2xlarge to ml.p4d.24xlarge) increases FLOPS and GPU memory bandwidth, allowing each epoch's forward/backward passes to complete faster. Option C is correct because SageMaker distributed training (e.g., data parallelism or model parallelism libraries) spreads the workload across multiple instances, so the dataset or model is processed in parallel and wall-clock training time is reduced. Option D is correct because Pipe input mode streams data directly from S3 to the training container instead of downloading the full dataset to local EBS first, eliminating the initial download bottleneck and improving throughput for large datasets. Option A is incorrect because decreasing the batch size means more iterations per epoch and typically slower overall training. Option E is incorrect because simply increasing the batch size does not guarantee reduced training time and can hurt convergence, requiring more epochs.

Answer

Decrease the batch size.

Answer

Increase the batch size.

Question 12

A data scientist is training a neural network for a multi-class classification problem. The model is overfitting. Which TWO of the following techniques can help reduce overfitting? (Choose two.)

Accepted Answer

Add dropout layers after the hidden layers.. Option B is correct because dropout layers randomly deactivate a fraction of neurons during each training iteration, which prevents the network from relying on specific neurons and forces it to learn more robust, generalizable features, directly reducing overfitting. Option D is correct because L2 regularization adds a penalty proportional to the squared magnitude of the weights to the loss function, discouraging large weights and thereby constraining model complexity to improve generalization. Option A is not correct because increasing hidden layers raises model capacity, which typically worsens overfitting rather than reducing it. Option C is not correct because decreasing the learning rate only affects the optimization step size and convergence speed, not the model's tendency to overfit. Option E is not correct because reducing batch size changes gradient noise and training dynamics but does not by itself serve as a regularization technique to reduce overfitting.

Answer

Increase the number of hidden layers.

Answer

Decrease the learning rate.

Answer

Reduce the batch size.

Question 13

A data engineer runs the above CLI command and sees that the bucket contains many small Parquet files (1 MB each) under the prefix. When querying this data with Athena, the query performance is poor and costs are high. Which approach would MOST improve performance and reduce cost?

Accepted Answer

Consolidate the small files into fewer, larger Parquet files. C is correct because consolidating many small Parquet files into fewer, larger files (e.g., 128–256 MB each) reduces the overhead of Amazon Athena's file listing and metadata operations, and improves compression and predicate pushdown efficiency. Parquet is a columnar format optimized for analytics, so keeping it while reducing file count directly addresses the root cause of poor performance and high cost.

Answer

Convert the files to JSON format

Answer

Convert the files to CSV format

Answer

Add more partitions by including hour in the prefix

Question 14

A company uses SageMaker to deploy a model for real-time inference. The model is a large ensemble that requires 8 GB of memory and has high latency. The team wants to reduce latency without increasing cost. Which strategy is most effective?

Accepted Answer

Use SageMaker Neo to compile the model for the target instance.. SageMaker Neo compiles and optimizes the model for the target hardware, reducing latency and memory footprint without increasing cost. Option A is wrong because a larger instance would increase cost. Option B is wrong because deploying on multiple instances increases cost and complexity without necessarily reducing latency. Option D is wrong because batch transform is for offline inference, not real-time, and does not address the low latency requirement.

Answer

Use a larger instance type with more memory.

Answer

Deploy the model on multiple instances behind a load balancer.

Answer

Switch from real-time inference to batch transform.

Question 15

A company needs to move 10 TB of data from an on-premises NAS to Amazon S3 over a 100 Mbps internet connection. The transfer must complete within 3 days. Which solution is the most appropriate?

Accepted Answer

Use AWS Snowball Edge to transfer the data. AWS Snowball Edge is a physical petabyte-scale data transfer device designed for scenarios where network bandwidth is insufficient to meet transfer deadlines. At 100 Mbps, transferring 10 TB would take approximately 10-11 days (10 TB = 80,000,000 Mb ÷ 100 Mbps ≈ 800,000 seconds ≈ 9.26 days at theoretical max, realistically longer), far exceeding the 3-day requirement. Snowball Edge bypasses the internet entirely by shipping data physically, making it the only viable option here.

Answer

Use AWS DataSync to transfer over the internet

Answer

Enable S3 Transfer Acceleration on the bucket

Answer

Use AWS CLI to copy data directly over the internet

Question 16

A company is building a recommendation system using matrix factorization. The training data contains user-item interactions. The model performs well on the training set but poorly on the test set. Which regularization technique should be applied to improve generalization?

Accepted Answer

Add L2 regularization to the user and item latent factors. L2 regularization (weight decay) penalizes large values in the user and item latent factor matrices, which helps prevent overfitting by encouraging the model to learn smoother, more generalizable representations. This is the standard regularization technique used in matrix factorization for collaborative filtering, as it directly controls the magnitude of the latent vectors without inducing sparsity.

Answer

Add L1 regularization to the user and item latent factors

Answer

Apply dropout to the latent factors during training

Answer

Use batch normalization on the factors

Question 17

Which TWO techniques can be used to detect multicollinearity among numerical features during exploratory data analysis? (Choose two.)

Accepted Answer

Compute a correlation matrix and look for pairs with absolute correlation > 0.8.. Option B is correct because computing a Pearson correlation matrix on numerical features and flagging pairs whose absolute correlation exceeds a threshold such as 0.8 is a standard, direct way to spot strong linear relationships that indicate multicollinearity. Option D is correct because the Variance Inflation Factor quantifies how much a feature's variance is inflated by its linear dependence on the other features, with VIF values above roughly 5 or 10 signaling problematic multicollinearity. Option A is not a detection technique per se: PCA loadings describe variance structure and dimensionality but do not measure collinearity between original features. Option C is wrong because a t-test compares means between groups, not the linear association between two continuous features. Option E is wrong because the chi-square test of independence applies to categorical variables, not numerical features.

Answer

Apply Principal Component Analysis (PCA) and examine loadings.

Answer

Perform a t-test between each pair of features.

Answer

Use a chi-square test of independence.

Question 18

A data scientist is training a deep learning model for image classification using Amazon SageMaker. The training job is taking too long. The data scientist notices that GPU utilization is low (around 30%). Which action is most likely to improve GPU utilization and reduce training time?

Accepted Answer

Increase the batch size. Low GPU utilization (around 30%) indicates that the GPU is spending too much time idle while waiting for data batches to be processed. Increasing the batch size allows each training step to process more samples per forward/backward pass, which increases computational load on the GPU and improves hardware utilization. This directly reduces the number of steps needed per epoch, thereby decreasing overall training time.

Answer

Use a smaller instance type

Answer

Increase the learning rate

Answer

Reduce the batch size

Question 19

Which THREE factors should be considered when choosing between Amazon Kinesis Data Streams and Amazon Kinesis Data Firehose for a real-time data ingestion pipeline? (Choose three.)

Accepted Answer

The need for custom real-time processing logic using consumer applications.. Option B is correct because Kinesis Data Streams is designed for custom consumer applications (using KCL, Kinesis Client Library, or AWS Lambda) that implement bespoke real-time processing logic, whereas Firehose only delivers to destinations with optional Lambda-based transformation. Option C is correct because latency is a key differentiator: Data Streams offers sub-second to low-second latency for consumers, while Firehose buffers data and delivers in micro-batches, typically resulting in 60-second-or-more latency (or as low as ~5 seconds with near-real-time settings). Option D is correct because Data Streams requires you to provision and manage shards (or use on-demand mode) and handle scaling/resizing, while Firehose is fully managed and auto-scales without shard management. Option A is not correct because built-in data transformation and analytics is not a primary differentiator in this comparison — Firehose offers Lambda-based transformation and Data Streams integrates with Kinesis Data Analytics, so both can support transformation/analytics to varying degrees. Option E is not correct because neither service natively guarantees exactly-once delivery semantics end-to-end; Data Streams provides at-least-once delivery with record-level ordering per shard, and Firehose provides at-least-once delivery to destinations, so exactly-once is not a valid selection factor between them.

Answer

The need for built-in data transformation and analytics.

Answer

The requirement for exactly-once delivery semantics.

Question 20

Refer to the exhibit. A SageMaker training job using the built-in Linear Learner algorithm fails with 'Loss function returned NaN'. Which hyperparameter change is MOST likely to resolve this issue?

Accepted Answer

Reduce learning_rate to 0.01. The 'Loss function returned NaN' error in SageMaker's built-in Linear Learner algorithm typically occurs when the learning rate is too high, causing gradient updates to overshoot optimal parameters and diverge. Reducing the learning rate to 0.01 stabilizes training by ensuring smaller, more controlled weight updates, preventing numerical instability that leads to NaN loss.

Answer

Increase the learning rate to 0.5

Answer

Increase mini_batch_size to 2000

Answer

Decrease epochs to 5