AZ MLO flashcards
AZ MLO · Build recall with a focused review session.
32 cards. Read the term and the definition here. The deck at the top of this page is optional.
- Workspace
- Top-level Azure Machine Learning resource that contains compute, datastores, datasets, models, endpoints, and experiments; provides a central place to manage all ML assets.
- Datastore
- A registered connection to Azure storage (Blob, Data Lake, SQL, file share) that lets you reference data by name without embedding credentials in code.
- Data asset
- A versioned reference to data in a datastore; can be a URI (file/folder) or MLTable (schema-aware) and is used as input to jobs and pipelines.
- Compute cluster
- A scalable set of nodes used for training jobs; supports CPU or GPU, autoscaling, and low-priority nodes for cost savings on fault-tolerant workloads.
- Environment
- A reproducible execution context defined by a Docker image and conda YAML; registered in the workspace and referenced by jobs and deployments.
- Managed identity
- An Azure AD identity automatically managed by Azure; used by the workspace to access storage and other resources without storing credentials.
- Command job
- An Azure ML job that runs a script or command on a compute target; defined in YAML with inputs, outputs, environment, and compute.
- Data leakage
- When information from the target or test set is inadvertently used during training; prevented by splitting data before preprocessing that learns parameters.
- MLTable
- A schema-aware data asset type in Azure ML that supports delimited files and Parquet; enables type conversion and column selection at read time.
- AutoML
- Automated machine learning that tries multiple algorithms and featurization steps; configured with task, primary metric, timeout, and cross-validation folds.
- Sweep job
- A hyperparameter tuning job that searches a space using random, grid, or Bayesian sampling and can early-terminate poor runs with Bandit or Median stopping.
- MLflow tracking
- Open-source logging API integrated with Azure ML; log metrics, parameters, and artifacts from training scripts for experiment comparison and lineage.
- Cross-validation
- A resampling technique that splits data into k folds, trains on k-1 and validates on the remaining fold, then averages metrics for robust evaluation.
- Fairlearn
- A Python package for assessing and mitigating fairness; computes metrics like demographic parity difference and equalized odds difference across sensitive groups.
- Error analysis
- A Responsible AI tool that identifies cohorts (subgroups) with high error rates to guide data collection and model improvement.
- Distributed training
- Training a model across multiple nodes or GPUs using frameworks like PyTorch DDP, Horovod, or MPI; configured via instance count and process count per node.
- Model registration
- Storing a model in the workspace registry with a name and version; enables deployment, lineage tracking, and lifecycle management.
- Online endpoint
- A real-time deployment target for low-latency scoring; supports autoscaling, traffic splitting across deployments, and key or token authentication.
- Batch endpoint
- A deployment target for large-scale asynchronous scoring; invoked with a data asset input and output location, runs on a compute cluster.
- Scoring script
- A Python file with init() to load the model and run() to process raw JSON input and return predictions; required for online and batch deployments.
- Autoscaling
- Automatically adjusting the number of endpoint instances based on CPU, memory, or custom metrics; configured with min and max instance counts.
- CI/CD pipeline
- Automated build and release process using Azure DevOps or GitHub Actions with Azure ML CLI (v2) to register models and deploy endpoints from YAML definitions.
- Private endpoint
- A network interface that connects privately to an Azure ML workspace or endpoint; used to secure production deployments and disable public access.
- Data drift
- A change in the distribution of production input data compared to a reference dataset; detected using metrics like population stability index or Jensen-Shannon distance.
- Prediction drift
- A shift in the distribution of model outputs over time; monitored on online endpoints to detect degradation and trigger retraining.
- Azure Monitor
- Azure service for collecting endpoint telemetry such as request count, latency, and errors; integrated with Application Insights for detailed diagnostics.
- Model registry
- A catalog of model versions with tags (e.g., staging, production) used to manage lifecycle, compare versions, and promote models to deployment.
- Retraining pipeline
- An automated pipeline triggered by drift alerts or schedules that retrains a model, registers a new version, and compares performance before promotion.
- Experiment
- A logical grouping of runs in Azure ML; used to track, compare, and filter job results by metrics, tags, and properties.
- SHAP
- A game-theoretic method for explaining model predictions; provides global feature importance and local explanations for individual predictions.
- Model card
- A document describing a model's purpose, training data, evaluation metrics, and limitations; stored with the registered model for transparency and auditability.
- Differential privacy
- A technique that adds calibrated noise to data or queries to protect individual privacy while allowing aggregate analysis; implemented via libraries like SmartNoise.