Vertex AI ML Workflow

Data Preparation

1. Upload data
   • Check data requirements
   • Label data—manually, or pay for Google to do this via the Vertex AI console
   • Upload—local, BigQuery or Cloud Storage
2. Feature engineers
   • Data prep before training—time-consuming and tedious
   • Vertex AI Feature Store
     • Organize, store and server ML features
     • Engineers can use features available in Feature Store dictionary to build a dataset
     • Benefits: features shareable, reusable and scalable, easy to use

Model Training

• Supervised learning—task driven, identifies a goal
  • Classification—predict a category
  • Regression—predict a continuous number
• Unsupervised learning—data-driven, identifies a pattern
  • Clustering
  • Association—identify underlying relationships
  • Dimensionality reduction—transform data from a high-dimensional space to a low-dimensional space
• Hyperparameters—user-defined training parameters

Model Evaluation

• Confusion Matrix
  • Precision and Recall
• Feature importance
  • Bar chart shown in Vertex AI showing which features are important to the model—explainable AI

Model Deployment and Monitoring

• Model serving
  1. Model deployment
  2. Model monitoring

MLOps

• Solves pain point—data and code continuously evolving
• MLOps ~ DevOps
  • Continuous integration
  • Continuous training
  • Continuous delivery
• Automate serving of model
• Monitoring at every step of ML workflow

Model Deployment

• Deploy to endpoint—when immediate results and low latency required, e.g. recommendations
• Batch prediction—no immediate response required
• Offline prediction—deploy outside of Google Cloud

Monitoring

• Vertex AI Pipelines
  • Automates, monitors, governs ML systems
  • Orchestrates workflow
  • Serverless
  • Triggers alerts