ML Pipeline Day 3

You are preparing a large set of CSV data for a training job using K-Means. Which of the following are NOT actions that you should expect to take in this scenario?

• Decide on the number of clusters you want to target.

• Use a mean or median strategy to populate any missing label data.

• Ensure that your IAM role has the iam:PassRole action.

• Convert the data to protobuf RecordIO format.

• Decide on the value you want to assign to k.

You are working with a machine learning team training an image classification model using MXNet on Amazon SageMaker. The requirements state that the model should be at least 85% accurate. The data appears to be of good quality, but the accuracy is around 48% during training with the test data. Most of the time wrong labels are being predicted. What should be done to help increase the accuracy of the model?

• Use Amazon SageMaker's automatic model tuning. Take the best performing hyperparameters and run multiple training jobs in parallel using Apache Spark and Spark ML

• Use Amazon SageMaker's automatic model tuning. Specify the objective metric and take the best performing parameters suggested by the service to use when training the model

• Use Amazon SageMaker's automatic model tuning. Use AWS Batch to run multiple batches of the training data with different hyper parameters specified during the autotuning job.

• Use Amazon SageMaker's automatic model tuning. Take the best performing hyperparameters and manually adjust them to meet your requirements.

You are a machine learning engineer on a team that is designing a deep sea exploration robot which will explore areas never before visited. The robot will be so deep that it's impractical to send commands for every motion. Rather then robot needs to know how to navigate among obstacles itself while collecting data. How would you frame this problem?

• This problem is not suitable for machine learning.

• Supervised Learning Problem

• Reinforcement Learning Problem

• Unsupervised Learning Problem

You have been working on a machine learning model for several iterations and feel that it is ready for production and allow real users to begin making inferences to it. You want to ensure that the models are ran on multiple instances in different availability zones. What steps can you take to ensure this occurs?

• Use Amazon SageMaker hosting services and specify a single instance. Use Route53 with failover routing policy to ensure users are routed to different availability zone if the instance becomes unreachable

• Use Amazon SageMaker hosting services, deploy two different variants of the model routing 50% of the traffic to one availability zone and the other 50% to the other availability zone

• Use Amazon SageMaker hosting services, specify two or more instances and specify multiple availability zones you want to launch models in

• Use Amazon SageMaker hosting services and specify two or more instances. Amazon SageMaker launches them in multiple availability zones automatically

You are training a model using a dataset with credit card numbers stored in Amazon S3. What should be done to ensure these credit cards are encrypted before and during model training?

• When calling the SageMaker SDK training job, ensure the SSE-KMS is used as a parameter during the creation of the training job.

• Create a Lambda function that is invoked when the training job starts to apply SSE-KMS key to the data before starting the training process.

• Ensure the S3 bucket and data have an SSE-KMS key associated with it, and specify the same SSE-KMS Key ID when you create the SageMaker notebook instance and training job.

• Create a SageMaker notebook instance with an SSE-KMS key associated with it. After loading the S3 data onto the notebook instance, encrypt it using SSE-KMS before feeding it into the training job.