AI and ML Fundamentals #1
SageMaker Pipelines automating machine learning workflows
Remember the correct answer, 'SageMaker Pipelines'.
・It defines and runs steps such as data preparation, training, evaluation, and deployment as an automated and reproducible workflow.
・Because the steps can be run the same way every time, it is the core of MLOps.
Feature Store (storing features), Data Wrangler (preprocessing), and JumpStart (model hub) are not pipelines that automate the entire workflow; they each relate to individual steps or resources.
Classification, regression, clustering, and dimensionality reduction
Remember the differences among the four representative machine learning tasks.
・Classification: supervised learning that trains on labeled data and assigns an input to a category (discrete label) (for example, lesion / no lesion, dog / cat).
・Regression: also supervised learning, but it predicts a continuous number such as sales or temperature.
・Clustering: unsupervised learning that groups similar data without using labels (for example, customer segmentation).
・Dimensionality reduction: unsupervised learning that compresses features without using labels to make them easier to handle (used for visualization and preprocessing).
Identify them by output: category → classification, number → regression, unlabeled grouping → clustering, compression → dimensionality reduction.
Early stopping halting training before overfitting sets in
Remember the correct answer, 'early stopping'.
・It halts training at the point when validation accuracy turns toward worsening.
・By stopping before overfitting advances, it curbs overfitting.
'Reducing the data' and 'training for longer' both worsen overfitting, and 'lowering latency' is unrelated to curbing overfitting.
Overfitting: a model that fits the training data too closely
Remember the correct answer, 'overfitting'.
・A state where the model fits the training data excessively and memorizes noise and fine details.
・Even with high accuracy on the training data, accuracy drops on unseen data and the model cannot generalize.
Underfitting is a different state where accuracy is low even on the training data, and regularization (a curbing technique) and hyperparameters (configuration values) are not names of 'states' in the first place.
The machine learning lifecycle from data collection to operations
Remember the machine learning lifecycle phases and the work done in each.
・Data collection: gather the data to be used for training.
・Data preparation: get the data into shape through cleansing (imputing missing values, handling outliers, standardizing formats) and feature engineering (creating useful input variables).
・Model training: train the model with an algorithm and tune the hyperparameters.
・Evaluation: check model performance with metrics such as accuracy.
・Deployment: deploy the model to the production environment.
・Monitoring (operations): monitor the model's behavior and accuracy in production and serve inference results via an API.
In this question, 'data cleansing and feature engineering' belong to the data preparation phase. Production monitoring and inference API serving belong to the post-deployment operations phase, and hyperparameter tuning belongs to the model training phase.