MLS-C01 Amazon Web Services AWS Certified Machine Learning - Specialty Free Practice Exam Questions (2025 Updated)

The solution D is the best data visualization approach to determine the optimal value of k for the k-means clustering algorithm. The solution D involves the following steps:

Run the k-means clustering algorithm for a range of k. For each value of k, calculate the sum of squared errors (SSE). The SSE is a measure of how well the clusters fit the data. It is calculated by summing the squared distances of each data point to its closest cluster center. A lower SSE indicates a better fit, but it will always decrease as the number of clusters increases. Therefore, the goal is to find the smallest value of k that still has a low SSE1.

Plot a line chart of the SSE for each value of k. The line chart will show how the SSE changes as the value of k increases. Typically, the line chart will have a shape of an elbow, where the SSE drops rapidly at first and then levels off. The optimal value of k is the point after which the curve starts decreasing in a linear fashion. This point is also known as the elbow point, and it represents the balance between the number of clusters and the SSE1.

The other options are not suitable because:

Option A: Calculating the principal component analysis (PCA) components, running the k-means clustering algorithm for a range of k by using only the first two PCA components, and creating a scatter plot with a different color for each cluster will not accurately determine the optimal value of k. PCA is a technique that reduces the dimensionality of the data by transforming it into a new set of features that capture the most variance in the data. However, PCA may not preserve the original structure and distances of the data, and it may lose some information in the process. Therefore, running the k-means clustering algorithm on the PCA components may not reflect the true clusters in the data. Moreover, using only the first two PCA components may not capture enough variance to represent the data well. Furthermore, creating a scatter plot may not be reliable, as it depends on the subjective judgment of the data scientist to decide when the clusters look reasonably separated2.

Option B: Calculating the PCA components and creating a line plot of the number of components against the explained variance will not determine the optimal value of k. This approach is used to determine the optimal number of PCA components to use for dimensionality reduction, not for clustering. The explained variance is the ratio of the variance of each PCA component to the total variance of the data. The optimal number of PCA components is the point where adding more components does not significantly increase the explained variance. However, this number may not correspond to the optimal number of clusters, as PCA and k-means clustering have different objectives and assumptions2.

Option C: Creating a t-distributed stochastic neighbor embedding (t-SNE) plot for a range of perplexity values will not determine the optimal value of k. t-SNE is a technique that reduces the dimensionality of the data by embedding it into a lower-dimensional space, such as a two-dimensional plane. t-SNE preserves the local structure and distances of the data, and it can reveal clusters and patterns in the data. However, t-SNE does not assign labels or centroids to the clusters, and it does not provide a measure of how well the clusters fit the data. Therefore, t-SNE cannot determine the optimal number of clusters, as it only visualizes the data. Moreover, t-SNE depends on the perplexity parameter, which is a measure of how many neighbors each point considers. The perplexity parameter can affect the shape and size of the clusters, and there is no optimal value for it. Therefore, creating a t-SNE plot for a range of perplexity values may not be consistent or reliable3.

1: How to Determine the Optimal K for K-Means?

2: Principal Component Analysis

3: t-Distributed Stochastic Neighbor Embedding

The best solution for this scenario is to use Amazon Kinesis Data Streams to export the data from Amazon DynamoDB to Amazon S3, and then configure QuickSight to access the data in Amazon S3. This solution has the following advantages:

It allows near real-time data ingestion from DynamoDB to S3 using Kinesis Data Streams, which can capture and process data continuously and at scale1.

It enables QuickSight to access the data in S3 using the Athena connector, which supports federated queries to multiple data sources, including Kinesis Data Streams2.

It avoids the need to create and manage a Lambda function or a Glue crawler, which are required for the other solutions.

The other solutions have the following drawbacks:

Using AWS Glue to export the data from DynamoDB to S3 introduces additional latency and complexity, as Glue is a batch-oriented service that requires scheduling and configuration3.

Using an API call from QuickSight to access the data in DynamoDB directly is not possible, as QuickSight does not support direct querying of DynamoDB4.

Using Kinesis Data Firehose to export the data from DynamoDB to S3 is less efficient and flexible than using Kinesis Data Streams, as Firehose does not support custom data processing or transformation, and has a minimum buffer interval of 60 seconds5.

1: Amazon Kinesis Data Streams - Amazon Web Services

2: Visualize Amazon DynamoDB insights in Amazon QuickSight using the Amazon Athena DynamoDB connector and AWS Glue | AWS Big Data Blog

3: AWS Glue - Amazon Web Services

4: Visualising your Amazon DynamoDB data with Amazon QuickSight - DEV Community

5: Amazon Kinesis Data Firehose - Amazon Web Services

The best solution for this scenario is to use SageMaker Clarify to generate the explanation report and attach it to the predicted results. SageMaker Clarify provides tools to help explain how machine learning (ML) models make predictions using a model-agnostic feature attribution approach based on SHAP values. It can also detect and measure potential bias in the data and the model. SageMaker Clarify can generate explanation reports during data preparation, model training, and model deployment. The reports include metrics, graphs, and examples that help understand the model behavior and predictions. The reports can be attached to the predicted results using the SageMaker SDK or the SageMaker API.

The other solutions are less optimal because they require more development effort and additional services. Using SageMaker Model Debugger would require modifying the training script to save the model output tensors and writing custom rules to debug and explain the predictions. Using AWS Lambda would require writing code to invoke the ML model, compute the feature importance and partial dependence plots, and generate and attach the explanation report. Using custom Amazon CloudWatch metrics would require writing code to publish the metrics, create dashboards, and generate and attach the explanation report.

Bias Detection and Model Explainability - Amazon SageMaker Clarify - AWS

Amazon SageMaker Clarify Model Explainability

Amazon SageMaker Clarify: Machine Learning Bias Detection and Explainability

GitHub - aws/amazon-sagemaker-clarify: Fairness Aware Machine Learning

 The engineer should create a private workforce and manifest file, and then create a labeling job by using the built-in bounding box task type in Amazon SageMaker Ground Truth. This will allow the engineer to build the labeling pipeline with the least effort.

A private workforce is a group of workers that you manage and who have access to your labeling tasks. You can use a private workforce to label sensitive data that requires confidentiality, such as medical images. You can create a private workforce by using Amazon Cognito and inviting workers by email. You can also use AWS Single Sign-On or your own authentication system to manage your private workforce.

A manifest file is a JSON file that lists the Amazon S3 locations of your input data. You can use a manifest file to specify the data objects that you want to label in your labeling job. You can create a manifest file by using the AWS CLI, the AWS SDK, or the Amazon SageMaker console.

A labeling job is a process that sends your input data to workers for labeling. You can use the Amazon SageMaker console to create a labeling job and choose from several built-in task types, such as image classification, text classification, semantic segmentation, and bounding box. A bounding box task type allows workers to draw boxes around objects in an image and assign labels to them. This is suitable for object detection tasks, such as identifying areas of concern on CT scans.

Create and Manage Workforces - Amazon SageMaker

Use Input and Output Data - Amazon SageMaker

Create a Labeling Job - Amazon SageMaker

Bounding Box Task Type - Amazon SageMaker

The true class frequency for Romance is the percentage of movies that are actually Romance out of all the movies. This can be calculated by dividing the sum of the true values for Romance by the total number of movies. The predicted class frequency for Adventure is the percentage of movies that are predicted to be Adventure out of all the movies. This can be calculated by dividing the sum of the predicted values for Adventure by the total number of movies. Based on the confusion matrix, the true class frequency for Romance is 57.92% and the predicted class frequency for Adventure is 13.12%. References: Confusion Matrix, Classification Metrics

A convolutional neural network (CNN) is a type of deep learning model that can learn to extract features from images and perform tasks such as classification, segmentation, and detection1. ResNet is a popular CNN architecture that uses residual connections to overcome the problem of vanishing gradients and enable very deep networks2. For the task of predicting product quality based on sensor data, a CNN and ResNet approach can leverage the spatial structure of the data and learn complex patterns that distinguish different quality levels.

Convolutional Neural Networks (CNNs / ConvNets)

PyTorch ResNet: The Basics and a Quick Tutorial

 Quantile binning is a data transformation technique that can be used to handle skewed and non-linear numerical features. It divides the range of a feature into equal-sized bins based on the percentiles of the data. Each bin is assigned a numerical value that represents the midpoint of the bin. This way, the feature values are transformed into a more uniform distribution that can improve the performance of linear models. Quantile binning can also reduce the impact of outliers and noise in the data.

One-hot encoding, Cartesian product transformation, and normalization are not suitable for this scenario. One-hot encoding is used to transform categorical features into binary features. Cartesian product transformation is used to create new features by combining existing features. Normalization is used to scale numerical features to a standard range, but it does not change the shape of the distribution. References:

Data Transformations for Machine Learning

Quantile Binning Transformation

Serverless inference in SageMaker is designed for workloads with intermittent traffic and unpredictable usage patterns, which aligns with the media company's periodic high-traffic windows. Because the payload is less than 4 MB (serverless inference supports payloads up to 4 MB), serverless inference is appropriate and provisioned concurrency ensures the endpoints are warm and ready during peak times to minimize latency.

From AWS documentation:

“Amazon SageMaker Serverless Inference is ideal for applications with intermittent or unpredictable traffic. You can optionally enable provisioned concurrency to ensure that your endpoints are always ready to process requests during anticipated peak traffic hours.”

— AWS SageMaker Serverless Inference documentation

Using real-time inference with auto scaling (A) incurs a higher cost as it requires always-on instances. Batch transform (D) is for offline, large-scale inferences, not low-latency real-time inference. Asynchronous inference (C) is typically used for large payloads (>4 MB) or long processing times.

To prepare the data for Word2Vec, the Specialist needs to perform some preprocessing steps that can help reduce the noise and complexity of the data, as well as improve the quality of the embeddings. Some of the common preprocessing steps for Word2Vec are:

Normalizing all words by making the sentence lowercase: This can help reduce the vocabulary size and treat words with different capitalizations as the same word. For example, “Fox” and “fox” should be considered as the same word, not two different words.

Removing stop words using an English stopword dictionary: Stop words are words that are very common and do not carry much semantic meaning, such as “the”, “a”, “and”, etc. Removing them can help focus on the words that are more relevant and informative for the task.

Tokenizing the sentence into words: Tokenization is the process of splitting a sentence into smaller units, such as words or subwords. This is necessary for Word2Vec, as it operates on the word level and requires a list of words as input.

The other options are not necessary or appropriate for Word2Vec:

Performing part-of-speech tagging and keeping the action verb and the nouns only: Part-of-speech tagging is the process of assigning a grammatical category to each word, such as noun, verb, adjective, etc. This can be useful for some natural language processing tasks, but not for Word2Vec, as it can lose some important information and context by discarding other words.

Correcting the typography on “quck” to “quick”: Typo correction can be helpful for some tasks, but not for Word2Vec, as it can introduce errors and inconsistencies in the data. For example, if the typo is intentional or part of a dialect, correcting it can change the meaning or style of the sentence. Moreover, Word2Vec can learn to handle typos and variations in spelling by learning similar embeddings for them.

One-hot encoding all words in the sentence: One-hot encoding is a way of representing words as vectors of 0s and 1s, where only one element is 1 and the rest are 0. The index of the 1 element corresponds to the word’s position in the vocabulary. For example, if the vocabulary is [“cat”, “dog”, “fox”], then “cat” can be encoded as [1, 0, 0], “dog” as [0, 1, 0], and “fox” as [0, 0, 1]. This can be useful for some machine learning models, but not for Word2Vec, as it does not capture the semantic similarity and relationship between words. Word2Vec aims to learn dense and low-dimensional embeddings for words, where similar words have similar vectors.

The combination of steps that the ML engineer should take to train the model are to create a .lst file that contains a list of image files and corresponding class labels, upload the .lst file to Amazon S3, and initiate transfer learning by training the model using the images of less common species. This approach will allow the ML engineer to leverage the existing ImageNetV2 CNN model and fine-tune it with the new data using Pipe mode in SageMaker.

A .lst file is a text file that contains a list of image files and corresponding class labels, separated by tabs. The .lst file format is required for using the SageMaker image classification algorithm with Pipe mode. Pipe mode is a feature of SageMaker that enables streaming data directly from Amazon S3 to the training instances, without downloading the data first. Pipe mode can reduce the startup time, improve the I/O throughput, and enable training on large datasets that exceed the disk size limit. To use Pipe mode, the ML engineer needs to upload the .lst file to Amazon S3 and specify the S3 path as the input data channel for the training job1.

Transfer learning is a technique that enables reusing a pre-trained model for a new task by fine-tuning the model parameters with new data. Transfer learning can save time and computational resources, as well as improve the performance of the model, especially when the new task is similar to the original task. The SageMaker image classification algorithm supports transfer learning by allowing the ML engineer to specify the number of output classes and the number of layers to be retrained. The ML engineer can use the existing ImageNetV2 CNN model, which is trained on 1,000 classes of common objects, and fine-tune it with the new data of less common animal species, which is a similar task2.

The other options are either less effective or not supported by the SageMaker image classification algorithm. Using a ResNet model and initiating full training mode would require training the model from scratch, which would take more time and resources than transfer learning. Using an Inception model is not possible, as the SageMaker image classification algorithm only supports ResNet and ImageNetV2 models. Using an augmented manifest file in JSON Lines format is not compatible with Pipe mode, as Pipe mode only supports .lst files for image classification1.

1: Using Pipe input mode for Amazon SageMaker algorithms | AWS Machine Learning Blog

2: Image Classification Algorithm - Amazon SageMaker

The solution C meets the requirements because it minimizes costs for compute infrastructure, maximizes the scalability of resources for training, and facilitates the use of an ML model in low-connectivity environments. The solution C involves the following steps:

Move the training data to an Amazon S3 bucket. This will enable the company to store the large amount of data in a durable, scalable, and cost-effective way. It will also allow the company to access the data from the cloud for training and evaluation purposes1.

Train and evaluate the model by using Amazon SageMaker. This will enable the company to use a fully managed service that provides various features and tools for building, training, tuning, and deploying ML models. Amazon SageMaker can handle large-scale data processing and distributed training, and it can leverage the power of AWS compute resources such as Amazon EC2, Amazon EKS, and AWS Fargate2.

Optimize the model by using SageMaker Neo. This will enable the company to reduce the size of the model and improve its performance and efficiency. SageMaker Neo can compile the model into an executable that can run on various hardware platforms, such as CPUs, GPUs, and edge devices3.

Set up an edge device in the manufacturing facilities with AWS IoT Greengrass. This will enable the company to deploy the model on a local device that can run inference in real time, even in low-connectivity environments. AWS IoT Greengrass can extend AWS cloud capabilities to the edge, and it can securely communicate with the cloud for updates and synchronization4.

Deploy the model on the edge device. This will enable the company to automate quality control in its facilities by using the model to detect defects in new parts as they move on a conveyor belt. The model can run inference locally on the edge device without requiring internet connectivity, and it can send the results to the cloud when the connection is available4.

The other options are not suitable because:

Option A: Deploying the model on a SageMaker hosting services endpoint will not facilitate the use of the model in low-connectivity environments, as it will require internet access to perform inference. Moreover, it may incur higher costs for hosting and data transfer than deploying the model on an edge device.

Option B: Training and evaluating the model on premises will not minimize costs for compute infrastructure, as it will require the company to maintain and upgrade its own hardware and software. Moreover, it will not maximize the scalability of resources for training, as it will limit the company’s ability to leverage the cloud’s elasticity and flexibility.

Option D: Training the model on premises will not minimize costs for compute infrastructure, nor maximize the scalability of resources for training, for the same reasons as option B.

1: Amazon S3

2: Amazon SageMaker

3: SageMaker Neo

4: AWS IoT Greengrass

DeepAR is a supervised algorithm that can leverage historical data from related time series to predict demand for new items, unlike ARIMA which requires sufficient historical data for the new item.

From AWS documentation:

“The DeepAR algorithm is a supervised learning algorithm for forecasting scalar (one-dimensional) time series. It can also forecast demand for new items by learning patterns from similar time series.”

— AWS SageMaker DeepAR documentation

 The best solution for building a line-counting application for use in a quick-service restaurant is to use the following steps:

Build a custom model in Amazon SageMaker to recognize the number of people in an image. Amazon SageMaker is a fully managed service that provides tools and workflows for building, training, and deploying machine learning models. A custom model can be tailored to the specific use case of line-counting and achieve higher accuracy than a generic model1

Deploy AWS DeepLens cameras in the restaurant to capture video. AWS DeepLens is a wireless video camera that integrates with Amazon SageMaker and AWS Lambda. It can run machine learning inference locally on the device without requiring internet connectivity or streaming video to the cloud. This reduces the bandwidth consumption and latency of the application2

Deploy the model to the cameras. AWS DeepLens allows users to deploy trained models from Amazon SageMaker to the cameras with a few clicks. The cameras can then use the model to process the video frames and count the number of people in each frame2

Deploy an AWS Lambda function to the cameras to use the model to count people and send an Amazon Simple Notification Service (Amazon SNS) notification if the line is too long. AWS Lambda is a serverless computing service that lets users run code without provisioning or managing servers. AWS DeepLens supports running Lambda functions on the device to perform actions based on the inference results. Amazon SNS is a service that enables users to send notifications to subscribers via email, SMS, or mobile push23

The other options are incorrect because they either require internet connectivity or streaming video to the cloud, which may impact the bandwidth and performance of the application. For example:

Option A uses Amazon Kinesis Video Streams to stream the data to AWS over the restaurant’s existing internet connection. Amazon Kinesis Video Streams is a service that enables users to capture, process, and store video streams for analytics and machine learning. However, this option requires streaming multiple video streams to the cloud, which may consume a lot of bandwidth and cause network congestion. It also requires internet connectivity, which may not be reliable or available in some locations4

Option B uses Amazon Rekognition on the AWS DeepLens device. Amazon Rekognition is a service that provides computer vision capabilities, such as face detection, face recognition, and object detection. However, this option requires calling the Amazon Rekognition API over the internet, which may introduce latency and require bandwidth. It also uses a generic face detection model, which may not be optimized for the line-counting use case.

Option C uses Amazon SageMaker to build a custom model and an Amazon SageMaker endpoint to call the model. Amazon SageMaker endpoints are hosted web services that allow users to perform inference on their models. However, this option requires sending the images to the endpoint over the internet, which may consume bandwidth and introduce latency. It also requires internet connectivity, which may not be reliable or available in some locations.

1: Amazon SageMaker – Machine Learning Service - AWS

2: AWS DeepLens - Deep learning enabled video camera - AWS

3: Amazon Simple Notification Service (SNS) - AWS

4: Amazon Kinesis Video Streams - Amazon Web Services

Amazon Rekognition – Video and Image - AWS

Deploy a Model - Amazon SageMaker

The Data Science team is facing a problem of imbalanced data, where the positive class (paid users) is much less frequent than the negative class (non-paid users). This can cause the random forest model to be biased towards the majority class and have poor performance on the minority class. To mitigate this issue, the Data Science team can try the following approaches:

C. Generate more positive samples by duplicating the positive samples and adding a small amount of noise to the duplicated data. This is a technique called data augmentation, which can help increase the size and diversity of the training data for the minority class. This can help the random forest model learn more features and patterns from the positive class and reduce the imbalance ratio.

D. Change the cost function so that false negatives have a higher impact on the cost value than false positives. This is a technique called cost-sensitive learning, which can assign different weights or costs to different classes or errors. By assigning a higher cost to false negatives (predicting non-paid when the user is actually paid), the random forest model can be more sensitive to the minority class and try to minimize the misclassification of the positive class.

Bagging and Random Forest for Imbalanced Classification

Surviving in a Random Forest with Imbalanced Datasets

machine learning - random forest for imbalanced data? - Cross Validated

Biased Random Forest For Dealing With the Class Imbalance Problem

The solution that the data scientist should use to improve the performance of the model is to apply the Synthetic Minority Oversampling Technique (SMOTE) on the minority class in the training dataset, and retrain the model with the updated training data. This solution can address the problem of class imbalance in the dataset, which can affect the model’s ability to learn from the rare but important positive class (fraud).

Class imbalance is a common issue in machine learning, especially for classification tasks. It occurs when one class (usually the positive or target class) is significantly underrepresented in the dataset compared to the other class (usually the negative or non-target class). For example, in the credit card fraud detection problem, the positive class (fraud) is much less frequent than the negative class (fair transactions). This can cause the model to be biased towards the majority class, and fail to capture the characteristics and patterns of the minority class. As a result, the model may have a high overall accuracy, but a low recall or true positive rate for the minority class, which means it misses many fraudulent transactions.

SMOTE is a technique that can help mitigate the class imbalance problem by generating synthetic samples for the minority class. SMOTE works by finding the k-nearest neighbors of each minority class instance, and randomly creating new instances along the line segments connecting them. This way, SMOTE can increase the number and diversity of the minority class instances, without duplicating or losing any information. By applying SMOTE on the minority class in the training dataset, the data scientist can balance the classes and improve the model’s performance on the positive class1.

The other options are either ineffective or counterproductive. Applying SMOTE on the majority class would not balance the classes, but increase the imbalance and the size of the dataset. Undersampling the minority class would reduce the number of instances available for the model to learn from, and potentially lose some important information. Oversampling the majority class would also increase the imbalance and the size of the dataset, and introduce redundancy and overfitting.

1: SMOTE for Imbalanced Classification with Python - Machine Learning Mastery

 The solution that will meet the requirements with the least development time is to use Amazon Kinesis Data Firehose to stream the reservation data directly to Amazon S3, detect high-demand outliers by using Amazon QuickSight ML Insights, and visualize the data in QuickSight. This solution does not require any custom development or ML domain expertise, as it leverages the built-in features of QuickSight ML Insights to automatically run anomaly detection and generate insights on the streaming data. QuickSight ML Insights can also create a visualization dashboard that automatically refreshes with the most recent data, and allows the data scientist to explore the outliers and their key drivers. References:

1: Simplify and automate anomaly detection in streaming data with Amazon Lookout for Metrics | AWS Machine Learning Blog

2: Detecting outliers with ML-powered anomaly detection - Amazon QuickSight

3: Real-time Outlier Detection Over Streaming Data - IEEE Xplore

4: Towards a deep learning-based outlier detection … - Journal of Big Data

The solution that will meet the requirements with the least development effort is to use Amazon SageMaker Feature Store to set feature groups for the current features that the ML models use, assign the required metadata for each feature, and use Amazon QuickSight to analyze the metadata. This solution can leverage the existing AWS services and features to perform feature-level metadata analysis and reporting.

Amazon SageMaker Feature Store is a fully managed, purpose-built repository to store, update, search, and share machine learning (ML) features. The service provides feature management capabilities such as enabling easy feature reuse, low latency serving, time travel, and ensuring consistency between features used in training and inference workflows. A feature group is a logical grouping of ML features whose organization and structure is defined by a feature group schema. A feature group schema consists of a list of feature definitions, each of which specifies the name, type, and metadata of a feature. The metadata can include information such as data sensitivity, authorship, description, and parameters. The metadata can help make features discoverable, understandable, and traceable. Amazon SageMaker Feature Store allows users to set feature groups for the current features that the ML models use, and assign the required metadata for each feature using the AWS SDK for Python (Boto3), AWS Command Line Interface (AWS CLI), or Amazon SageMaker Studio1.

Amazon QuickSight is a fully managed, serverless business intelligence service that makes it easy to create and publish interactive dashboards that include ML insights. Amazon QuickSight can connect to various data sources, such as Amazon S3, Amazon Athena, Amazon Redshift, and Amazon SageMaker Feature Store, and analyze the data using standard SQL or built-in ML-powered analytics. Amazon QuickSight can also create rich visualizations and reports that can be accessed from any device, and securely shared with anyone inside or outside an organization. Amazon QuickSight can be used to analyze the metadata of the features stored in Amazon SageMaker Feature Store, and generate a report that summarizes the metadata analysis2.

The other options are either more complex or less effective than the proposed solution. Using Amazon SageMaker Data Wrangler to select the features and create a data flow to perform feature-level metadata analysis would require additional steps and resources, and may not capture all the metadata attributes that the company requires. Creating an Amazon DynamoDB table to store feature-level metadata would introduce redundancy and inconsistency, as the metadata is already stored in Amazon SageMaker Feature Store. Using SageMaker Studio to analyze the metadata would not generate a report that can be easily shared and accessed by the company.

1: Amazon SageMaker Feature Store – Amazon Web Services

2: Amazon QuickSight – Business Intelligence Service - Amazon Web Services

Amazon Comprehend is a natural language processing (NLP) service that can perform part-of-speech tagging and key phrase extraction tasks. AWS Deep Learning Containers are Docker images that are pre-installed with popular deep learning frameworks such as Apache MXNet. Amazon SageMaker is a fully managed service that can help build, train, and deploy machine learning models. Using Amazon Comprehend for the text preprocessing tasks and AWS Deep Learning Containers with Amazon SageMaker to build the custom classifier is the solution that can be built most quickly to meet the requirements.

Amazon Comprehend

AWS Deep Learning Containers

Amazon SageMaker

 To improve the test error for the image classifier, the Machine Learning Specialist should use the technique of increasing the training data by adding variation in rotation for training images. This technique is called data augmentation, which is a way of artificially expanding the size and diversity of the training dataset by applying various transformations to the original images, such as rotation, flipping, cropping, scaling, etc. Data augmentation can help the model learn more robust features that are invariant to the orientation, position, and size of the objects in the images. This can improve the generalization ability of the model and reduce the test error, especially for cases where the images are not well-aligned or have different perspectives1.

1: Image Augmentation - Amazon SageMaker