AI Simplified: Introduction to Vertex AI

Overview

• Presenter: Priyanka Vergaria
• Goal: Journey from datasets to deployed machine learning (ML) models using Vertex AI.
• Importance: Leveraging data for predictions to improve apps and user experience.
• Target Audience: Teams with varying levels of ML expertise (novice to experts).

Introduction to Vertex AI

• Purpose: Accelerates AI innovation by catering to different expertise levels.
• Features: Provides tools for every step in the ML workflow for varied model types and levels of expertise.

Typical Machine Learning Workflow

1. Define Prediction Task
2. Ingest Data
3. Analyze and Transform Data
4. Create and Train the Model
5. Evaluate Model
6. Deploy Model for Predictions

Vertex AI Workflow Simplification

Data Preparation

• Ingestion, Analysis, and Transformation: Using managed datasets within Vertex AI.
• Tools: Create dataset by importing data through console or API.
• Labeling and Annotation: Directly within the console.

Model Training Options

• AutoML: Suitable for images, videos, text files, and tabular data. No need to write model code.
• Custom Models: For more control over model architecture (e.g., TensorFlow, PyTorch).

Model Assessment

• Evaluate and Optimize
• Explainable AI: Understand factors behind model predictions.

Deployment

• Deployment to an Endpoint: For online predictions via API or console.
• Scalable Resources: For low latency and scaled hardware.

Model Utilization

• Access Predictions: Command line interface, console UI, SDK, or APIs.

Dashboard Tour

Console Dashboard

• Overview: Recent datasets, models, and prediction tools.
• Left Panel: Detailed steps from datasets to predictions.

Datasets

• Creation: Based on data type and prediction task (image, tabular, text, video).
• Custom Models: For tasks outside predefined use cases.
• Dataset List: Shows created datasets.

Notebooks

• Customization: Create notebook instances with specific environments and GPUs.

Training

• Job Management: Create and manage training jobs.
• Methods: AutoML, AutoML Edge, Custom Training.
  • AutoML: Minimal effort, high-quality model.
  • AutoML Edge: Models optimized for edge devices.
  • Custom Training: Use any framework with pre-built or custom containers.
• Containers: Pre-built for TensorFlow, PyTorch, Scikit-learn, XGBoost; custom containers via Docker.
• Acceleration: Training with GPUs and hyperparameter tuning.

Models

• Model Management: View and import models for online/batch predictions.
• Endpoints: Create endpoints for serving online predictions.
• Scaling: Auto-scaling resources based on traffic, traffic splitting, logging.

Predictions

• Methods: UI, SDK for batch predictions from cloud storage.

Conclusion

• Vertex AI Tools: Supports entire ML workflow from data management to predictions.
• Future Episodes: Detailed exploration of each step in the ML workflow.
• Audience Interaction: Comments and discussions on individual ML use cases and workflows.