AWS ML Associate

Question 1

Performance metric: Measure the imbalance of positive outcomes between different facet values.

Answer 1

Difference in proportions of labels (DPL)

Question 2

Performance metric: Identify the difference in the predicted outcome as an input feature changes.

Answer 2

Partial dependence plots (PDPs)

Question 3

Performance metric: Quantify the contribution of each feature in a prediction.

Answer 3

Shapley values

Question 4

What should you use for data processing if it involves Tensorflow or Pytorch?

Answer 4

SageMaker

Question 5

What is the simplest way to prevent internet and data access to inference containers?

Answer 5

Sagemaker network isolation mode

Question 6

Create a baseline to monitor a Sagemaker model’s bias drift. For instance, you want it to weigh personal income over credit history for loan approval. How do you do this?

Answer 6

Create a SHAP baseline using the ‘ModelExplainabilityMonitor’ class. Generate a feature attribution baseline which will trigger when the observed feature attribution occurs.

Question 7

tool used to check for bias and explainability in datasets and models

Answer 7

SageMaker Clarify

Question 8

used to visualize and analyze intermediate tensors. Identify specific poor classifications in a CNN and make adjustments to improve model performance.

Answer 8

SageMaker with TensorBoard

Question 9

How do you strip PII from text-based user interactions

Answer 9

Amazon Comprehend

Question 10

RNN training: Exploding gradients causing a convergence issue. What feature can help address this issue?

Answer 10

Sagemaker Training Compiler. Optomises DL models to accelerate training by more efficiently using ML GPU instances.

Question 11

What instance types are supported by AWS Neuron SDKs for real-time inference on streaming video?

Answer 11

Inferentia instances (Inf2 family)

Question 12

What are used to centralize and standardize model documentation.

Answer 12

SageMaker Model Cards

Question 13

SageMaker Serverless Inference: What is the biggest consideration when deciding whether to use provisioned concurrency?

Answer 13

low-latency (avoiding cold-starts)

Question 14

(CloudWatch) What feature in the Logs Insights page is helpful in finding infrastructure monitoring through-lines in your query results?

Answer 14

The Patterns tab

Question 15

What is the primary purpose of Capacity Blocks for machine learning (ML)?

Answer 15

Reserve GPU instances for short-duration machine learning workloads on a future date.

Question 16

When using an embedded question to query a vector database for RAG, what should be returned?

Answer 16

The full text - not embeddings - of the nearest neighbor documents to enhance the query

Question 17

How can you use SageMaker Model Monitor to re-train your model?

Answer 17

Enable Data Capture, and use that data to retrain the model.

Question 18

Exploratory data visualization that can be used to identify hidden patterns, (ralationship analysis) such as an increase in specific item purchases or periods of frequent transactions

Answer 18

Heat Map

Question 19

Exploratory data visualization that helps with distribution analysis by binning ranges of data

Answer 19

Histogram

Question 20

Which storage service should be used when you need concurrent access from multiple Amazon EC2 instances to a Windows File Server for distributed training of the model.

Answer 20

FSx

Question 21

Best way to improve Kinesis ingestion performance

Answer 21

Batching

Question 22

What type of variables should be on the x-axis of a bar chart?

Answer 22

Categorical variables

Question 23

Data format for ingesting streamed, unstructured data

Answer 23

JSON lines

Question 24

When to choose Lustre over EFS

Answer 24

Only when super high performace, high volumes and extremely low latency is the requirement. EFS is go-to for distributed training and S3 is go-to for general storage

Question 25

Which text feature engineering technique would categorize customer feedback as positive or negative?

Answer 25

N-gram

Question 26

How can you reduce the dimensionality of the data while retaining most of the variation?

Answer 26

Principle Component Analysis

Question 27

How can you save time by storing curated features that can be accessed to train new models.

Answer 27

SageMaker Feature Store

Question 28

When to use EMR over Glue?

Answer 28

real-time shit. Glue is only for batch ETL

Question 29

What ETL solution allows you to do anomaly detection on real-time data?

Answer 29

Apache Spark on EMR

Question 30

How do you address class imbalances in text-based datasets?

Answer 30

text-based data augmentation, like synonym replacement, or text paraphrasing

Question 31

How can you use AWS Glue Data Quality to assess data before training an ML model?

Answer 31

data validation rules

Question 32

What does a Class Imbalance metric of .9 mean?

Answer 32

It means that the advantaged group is overrepresented in the data, and they are HIGHLY advantaged.

Question 33

Which data formats do most Amazon SageMaker algorithms support for training?

Answer 33

CSV and RecordIO-protobuf

Question 34

What is used to compare the distribution of labels in your data to the expected proportions?

Answer 34

Difference in Proportion of Labels (DPL)

Question 35

Which built-in algorithm is used for text classification and Word2Vec?

Answer 35

BlazingText

Question 36

Which built-in algorithm is a great choice for a supervised text translation model?

Answer 36

Sequence-to-Sequence algorithm

Question 37

Which deep learning frameworks are supported in Amazon SageMaker reinforcement learning (RL)? (2)

Answer 37

TensorFlow and Apache MXNet

Question 38

Difference between Lex v2 and Kendra?

Answer 38

Lex is a chat bot, Kendra is natural language search.

Question 39

What approach is referred as script mode when using Amazon SageMaker?

Answer 39

Using pre-set framework and dependencies, but providing your own custom training algorithms.

Question 40

Training: What is the most appropriate data ingestion mode for a large data set of historical data?

Answer 40

Pipe mode.

File mode involves ingesting the whole set at once and will not be as performative as streaming with pipe mode. Fast File mode is used for sequential data

Question 41

The model has billions of parameters, and training it on a single GPU would be infeasible due to memory constraints. How fix?

Answer 41

Model parallelism

Question 42

Main Use-Cases for Trainium instances?

Answer 42

large language models and natural language processing (NLP) training

Question 43

What is it called when you use multiple models of different types and aggragate their predictions in a heterogeneous model group?

Answer 43

Stacking

Question 44

Which hyperparameter tuning method is best for finding optimum hyperparameter values with limited compute resources?

Answer 44

Hyperband

Question 45

CNN is failing to generalize well, although it performs well on training data. What method can be used to help it adapt to unforeseen patterns?

Answer 45

dropout