Things I got wrong - MLA

Question 1

How many variables’ distributions are shown in a histogram?

Answer 1

1

Question 2

What is JSONL?

Answer 2

JSON lines. Used to encode JSON objects with each in a separate line.

Question 3

What is Parquet?

Answer 3

A columnar data format which is well suited to batch data processing

Question 4

What is a scatter plot useful for?

Answer 4

Testing the correlation between two continuous variables

Question 5

What is the main use of Managed Service for Apache Kafka?

Answer 5

Transfer data between different applications and systems

Question 6

Can EFS be used outside of EC2 instances?

Answer 6

No

Question 7

What is FSx for Lustre good for with regards to Sagemaker?

Answer 7

Distributed training. Works well due to its high performance

Question 8

Is Ground Truth using humans or machines to do the labelling?

Answer 8

Both!

Question 9

Do you change data’s format first or filter it first?

Answer 9

You filter it first, then change the format

Question 10

What does the “difference in proportions of labels” metric tell you?

Answer 10

The difference in the proportions of positive and negative labels between 2 groups. For example, identify how much the predicted churn changes between different demographics to be aware of potential biases

Question 11

What does Cramer’s V tell you?

Answer 11

The association between two categorical variables

Question 12

What type of splitting is used when there is a date or time series?

Answer 12

Ordered splitting. This means that the model only has access to historical data, otherwise it wouldn’t make sense to ask it to predict for dates it already knows

Question 13

What does Spearman’s rank correlation tell you and when should it be used?

Answer 13

To find the strength and direction of the relationship between two monotonic variables. Doesn’t assume a distribution

Question 14

Why might SageMaker processing jobs be advantageous over Glue for ML workloads?

Answer 14

SM processing jobs integrate w/ ML libraries like Tensorflow

Question 15

What is the purpose of Sagemaker batch transform?

Answer 15

Allows you to perform inference on large datasets without having to maintain a persistent endpoint

Question 16

What is the maximum processing time of Sagemaker serverless?

Answer 16

60 seconds

Question 17

What is random cut forest used for?

Answer 17

Anomaly detection in high-dimensional data, such as time series or multivariate data

Question 18

Do VPC gateway endpoints allow data transfer across regions? What about interface endpoints?

Answer 18

Gateway: no, Interface: yes

Question 19

What is the purpose of Sagemaker IP insights?

Answer 19

Identify IPs that deviate from normal patterns through the analysis of historical logs

Question 20

Do S3 access points have authentication or access control?

Answer 20

No

Question 21

What is Sagemaker DeepAR algorithm used for?

Answer 21

Time series forecasting only

Question 22

What type of instances are often good for deep learning inference?

Answer 22

Accelerated computing instances w/ GPUs or FGPAs

Question 23

What is the decorator that you have to use to integrate a custom step into a SageMaker Pipeline?

Answer 23

@step

Question 24

Can Comprehend be used to redact/anonymise PII? When would you use this over Macie?

Answer 24

Yes, and you’d use it for real-time or near-real-time use cases e.g. customer communications

Question 25

What is the difference in SageMaker Clarify and SageMaker Model Monitor?

Answer 25

SageMaker Clarify is explainability and bias detection. SageMaker Model Monitor is drift, model performance, managing costs

Question 26

Is using a notebook for production deployments efficient?

Answer 26

Not really, notebooks are better for getting a feel on how the system works and how you can interact with it

Question 27

What is model bias?

Answer 27

The idea that the output is being biased to a specific outcome for a specific group, e.g. old people always get bigger loans

Question 28

What is the feature attribution?

Answer 28

The idea that you can say which metric from the input data caused a certain outcome in the model's results, e.g. you can see that the 'income' field is contributing the most dominantly to the inferences of the model

Question 29

Where is pipe mode better over fast file mode?

Answer 29

Pipe mode is better if the data should be loaded sequentially, but otherwise fast file is probably more suitable

Question 30

Is Amazon MWAA serverless?

Answer 30

No

Question 31

What ensemble method should you use if you want to assemble multiple different algorithms/model types?

Answer 31

Stacking. Bagging and boosting typically work with one model type at a time.

Question 32

What is feature splitting?

Answer 32

Breaking a complex input feature into 2 smaller and simpler sub features.

Question 33

What is one hot encoding?

Answer 33

A technique to convert categorical variables into numerical ones

Question 34

What is the purpose of AWS X-Ray?

Answer 34

Finding bottlenecks and tracking the flow of execution in microservices architectures

Question 35

The ROC is the plot of the XXX XXX vs XXX XXX rates at various thresholds.

Answer 35

The ROC is the plot of the true positive vs false positive rates at various thresholds.

Question 36

What is ordinal encoding?

Answer 36

A technique used to encode categorical data into numerical data