Modeling Study Guide — MLS-C01

Modeling (36%) – AWS MLS-C01 Study Guide

This domain tests your ability to select the right algorithm for a problem, configure and launch SageMaker training jobs, tune hyperparameters, evaluate model performance, and apply explainability and bias detection tools. Expect scenario-based questions where you must justify an algorithm choice, diagnose a training problem, or pick the correct SageMaker feature for a given constraint.

SageMaker Built-in Algorithms: Tabular and Time-Series

Know the input format, use case, and key hyperparameters for each algorithm cold.

Critical distinctions:

• Use RCF for anomaly detection on unlabeled data; do not confuse with Isolation Forest (not a built-in).
• DeepAR requires related time-series in the same dataset; it outperforms ARIMA when you have many correlated series.
• Linear Learner supports both L1 and L2 regularization and automatically tunes the learning rate—use it as a fast, interpretable baseline before XGBoost.
• BlazingText in supervised mode is essentially FastText; in Word2Vec mode it is unsupervised.

SageMaker Built-in Algorithms: Computer Vision

• Image Classification: Assigns a single label per image. Uses ResNet. Supports full training and transfer learning (set use_pretrained_model=1).
• Object Detection: Bounding boxes + labels. Based on SSD with VGG or ResNet backbone.
• Semantic Segmentation: Pixel-level classification. Highest compute cost of the three; requires GPU instances.

All three accept RecordIO or augmented manifest files (S3-backed). For large datasets, augmented manifests avoid copying data—a common exam scenario about cost/efficiency.

SageMaker Built-in NLP Algorithms

• Sequence-to-Sequence (Seq2Seq): Machine translation, text summarization. Requires GPU. Input is tokenized text in RecordIO-Protobuf format.
• LDA (Latent Dirichlet Allocation): Unsupervised topic modeling. Use when you need interpretable topics from a document corpus.
• Neural Topic Model (NTM): Also topic modeling but neural-based. Faster to train than LDA; less interpretable. Choose NTM when scale matters, LDA when interpretability matters.

Custom Training Containers

SageMaker supports Script Mode (bring your own script with a managed container) and BYOC (bring your own container).

• Script Mode: Use pre-built TensorFlow, PyTorch, or MXNet containers. Pass your script via entry_point. This is the preferred path and the exam's default assumption.
• BYOC: Required when you need a custom runtime, specific library version, or non-supported framework. Container must implement /opt/ml/ directory conventions and respond to train entrypoint.
• Distributed training: Use SageMaker Data Parallelism Library for data-parallel workloads (built into PyTorch/TF containers). Use Model Parallelism Library when the model itself won't fit in a single GPU.