AWS Machine Learning Specialty (AWS) flashcards

31 cards. Read the term and the definition here. The deck at the top of this page is optional.

S3 training channel

The input data source for a SageMaker training job. The job's IAM role must allow s3:GetObject on the bucket and prefix, and data is streamed or downloaded to the training container.

Glue Data Catalog

A managed metastore holding table definitions, schemas, and partition metadata. Athena, Redshift Spectrum, and SageMaker query it so schema definitions stay consistent across engines.

Parquet columnar format

A compressed columnar file format supporting predicate pushdown. Because Athena bills per byte scanned, converting CSV to Parquet usually lowers query cost and improves performance.

Partition pruning

Skipping irrelevant data files by filtering on partition columns such as date or region. It reduces bytes scanned, but thousands of tiny partitions add per-partition overhead that hurts performance.

Kinesis Data Firehose

A managed delivery service that buffers streaming records and writes them to S3, Redshift, or OpenSearch. It is not a general-purpose consumer API, so custom processing needs Kinesis Data Streams or Lambda.

DMS change data capture

Database Migration Service replicates ongoing changes from an operational database. Combined with a Glue job targeting S3, it builds a continuously updated data lake.

Feature Store online store

A low-latency store backed by DynamoDB that serves feature values to real-time inference. The offline store holds Parquet history in S3 for training and batch scoring.

Glue DataBrew

A no-code visual tool for profiling and transforming tabular data. It surfaces missing values, duplicates, and outliers, and its recipes can be exported to Glue jobs.

IQR outlier rule

A value is a common outlier candidate when below Q1 minus 1.5 times the interquartile range or above Q3 plus 1.5 times the IQR. The rule is robust to extreme values.

Pearson correlation limits

Pearson correlation captures only linear association. Anscombe's quartet shows datasets with identical means, variances, and correlations but very different shapes, so plots are required.

Precision-recall AUC

An area-under-curve metric computed from the precision-recall curve. It is more informative than ROC-AUC when the positive class is rare and negative examples dominate.

Variance inflation factor

A measure of how much a predictor's variance is inflated by correlation with other predictors. Values above roughly five to ten suggest multicollinearity worth addressing.

Yeo-Johnson transform

A power transformation that reduces skew and, unlike Box-Cox, accepts zero and negative values, making it applicable to features that are not strictly positive.

Target leakage

A feature that encodes information about the label that would not be available at prediction time. It produces unrealistically high validation scores that collapse in production.

SageMaker Clarify

A SageMaker capability that computes bias metrics such as class imbalance and difference in positive proportions, and generates SHAP-based feature attributions explaining model predictions.

Train-serve skew

A mismatch between how features are computed during training and during inference. Shared transformation code and Feature Store reduce the risk of this silent accuracy loss.

XGBoost algorithm

A gradient-boosted decision tree ensemble available as a SageMaker built-in algorithm. It handles non-linear interactions and missing values and requires tuning of max_depth, eta, and num_round.

L1 versus L2 regularization

L1 (Lasso) pushes some coefficients exactly to zero, performing feature selection. L2 (Ridge) shrinks coefficients smoothly without eliminating them. Elastic Net blends both.

Automatic Model Tuning

A SageMaker hyperparameter search service using Bayesian optimization by default. You define an objective metric and bound cost with max_jobs and max_parallel_jobs.

DeepAR algorithm

A SageMaker built-in recurrent neural network for probabilistic time-series forecasting. It learns across many related series and outputs prediction distributions rather than point estimates.

BlazingText algorithm

A SageMaker built-in algorithm for text classification and word-vector generation. It scales to large corpora and supports supervised and unsupervised modes.

scale_pos_weight

An XGBoost hyperparameter that upweights the positive class during training to counter class imbalance. An alternative is adjusting the decision threshold after training.

k-fold cross-validation

A resampling scheme that trains on k-1 folds and validates on the held-out fold, repeating k times. It gives a lower-variance estimate than a single holdout split at higher compute cost.

JumpStart pre-trained models

A catalog of pre-trained task and foundation models in SageMaker that can be deployed directly or fine-tuned on your data, useful when labeled examples are scarce.

Serverless Inference

A SageMaker endpoint option that scales to zero when idle and suits intermittent or unpredictable traffic. It trades cold-start latency for lower cost.

Batch Transform

A SageMaker job that scores a large dataset from S3 and writes predictions back to S3 without a persistent endpoint. You pay only for the job duration.

Model Monitor

A SageMaker service that compares live endpoint data against a baseline to detect data drift, model-quality drift, and bias drift, emitting violations to CloudWatch.

SageMaker Pipelines

A service that orchestrates processing, training, evaluation, and conditional steps as a directed acyclic graph, and can register models with an approval status before deployment.

SageMaker Debugger

A training-time tool that captures tensors and system metrics and applies built-in rules such as loss-not-decreasing or overtraining to flag problems before training completes.

Production variants A/B testing

Two or more model variants served behind one endpoint with assigned traffic weights. You shift traffic gradually and compare metrics before promoting the new model.

Provisioned throughput

A SageMaker endpoint setting that reserves capacity for a model, guaranteeing consistent latency and reducing cold starts. It is billed hourly whether or not requests arrive.