Things I got wrong - MLS

Question 1

Can KDF perform data format transformations?

Answer 1

Yes, e.g. JSON to Parquet

Question 2

What type of data is well suited to imputation by deep learning?

Answer 2

Categorical

Question 3

What can object2vec turn into embeddings?

Answer 3

Full sentences

Question 4

What is incremental training?

Answer 4

Slowly retraining your model with new data

Question 5

How does KDF perform data transformations?

Answer 5

With built in lambda functions

Question 6

What type of algorithm are support vector machines and what can they be used for?

Answer 6

A supervised ML algorithm that can be employed for both classification and regression

Question 7

Can K-Means be used for classification?

Answer 7

No, it is for clustering

Question 8

What type of learning is classification?

Answer 8

Supervised learning with labelled data

Question 9

If you are overfitting, should you use more or less training data?

Answer 9

More, to introduce a greater diversity in the training data and force the model to generalise more

Question 10

When should you use PCA in randomised mode?

Answer 10

For datasets with a large amount of observations and features as it can then use an approximation algorithm to run faster

Question 11

If a model has high specificity, it means that nearly all ____ ______ have been weeded out.

Answer 11

If a model has high specificity, it means that nearly all false positives have been weeded out

Question 12

What is the equation for tf-idf?

Answer 12

TF x IDF, where:
TF is the amount of times a word appears in a document / total words in the document.
IDF is the the log of the total number of documents / the amount of documents containing the specified word.

Question 13

Can Glue transform data into RecordIO-Protobuf?

Answer 13

No

Question 14

If you are underfitting, what 2 actions can you take?

Answer 14

Use more features
Remove regularisation

Question 15

What is the fastest command to move data from S3 to Redshift?

Answer 15

COPY, this is much faster than INSERT

Question 16

What algorithm should you use to forecast sales for a product that has not been seen before specifically, but for which there are similar existing products?

Answer 16

DeepAR

Question 17

What type of algorithm is DeepAR?

Answer 17

Supervised RNN

Question 18

Can you integrate your own scripts with XGBoost?

Answer 18

Yes

Question 19

What are two scaling techniques that can be used to reduce the effect of outliers in your data?

Answer 19

Robust standardisation and logarithm transformation

Question 20

Does SageMaker support resource based policies and service linked roles?

Answer 20

No

Question 21

Does training with a SageMaker built-in algorithm require an IAM role?

Answer 21

Yes

Question 22

What is the read/write of each KDS shard?

Answer 22

Read max is 2MB
Write max is 1MB

Question 23

What two conditions must CSVs going into SageMaker supervised learning algorithms meet?

Answer 23

They should not have a header record and the target variable should be in the first column

Question 24

Are InvokeEndpoint calls to SageMaker monitored by CloudTrail?

Answer 24

No

Question 25

What is a Poisson distribution?

Answer 25

A probability distribution that is used to show how many times an event is likely to occur over a specified period

Question 26

What part of a dataset is standardised or normalised?

Answer 26

Specific numerical features, NOT the whole thing

Question 27

What is the difference between standardisation and normalisation?

Answer 27

Standardisation moves the mean to 0, but the standard deviation remains the same. Normalisation makes the values all between 0 and 1.

Question 28

What is elastic inference?

Answer 28

A way to speed up the throughput and decrease the latency of getting real-time inferences from SageMaker deep learning models

Question 29

What is AWS Panorama?

Answer 29

Enables you to do computer vision at the edge through AWS devices. Works with existing camera networks

Answer 30

K-fold validation splits data into k equal parts, trains the model k times, each time using a different fold as validation data and the remaining k-1 folds as training data. The model's performance is averaged across all k iterations. Common value for k is 5 or 10.