Introduction to Generative AI

Instructor

• Dr. Gwendolyn Stripling
• Position: AI Technical Curriculum Developer at Google Cloud

Course Outline

• Define Generative AI
• Explain how Generative AI works
• Describe Generative AI model types
• Describe Generative AI applications

What is Artificial Intelligence (AI)?

• AI: Creation of intelligent agents that reason, learn and act autonomously
• Difference between AI and Machine Learning (ML):
  • AI: A broad discipline, like physics
  • ML: A subfield of AI; trains models to make predictions from new data

Machine Learning

• Supervised Learning:
  • Uses labeled data (with tags like names, numbers)
  • Example: Predicting tips based on bill amount and order type
• Unsupervised Learning:
  • Uses unlabeled data
  • Example: Clustering employees based on tenure and income
• Process: Inputs (X) are fed into the model, which predicts outputs (Y); goal is to minimize error between predicted and actual values

Deep Learning

• Subset of machine learning
• Uses artificial neural networks to process complex patterns
• Artificial Neural Networks:
  • Inspired by the human brain
  • Made up of interconnected nodes (neurons)
  • Can use both labeled and unlabeled data (semi-supervised learning)

Generative AI (GenAI)

• Subset of deep learning
• Uses artificial neural networks
• Processes both labeled and unlabeled data
• Can use supervised, unsupervised, and semi-supervised methods

Model Types

• Discriminative Models:
  • Used to classify or predict labels for data points
  • Example: Classifying an image as a dog or not
• Generative Models:
  • Generate new data instances based on learned data
  • Example: Generating a picture of a dog

Key Distinctions

• Not GenAI: When output is a numerical value (e.g., predicted sales)
• GenAI: When output is natural language, image, audio, etc.

Process

• Takes training code, labeled and unlabeled data
• Builds a foundation model that can generate new content

Foundation Models

• Large AI models pre-trained on vast quantities of data
• Can be adapted to a wide range of downstream tasks
• Examples: Sentiment analysis, object recognition
• Applications:
  • Fraud detection, personalized customer support

Example Use Cases

• Code Generation: Convert Python to JSON, debug code, write documentation
• Tools:
  • Google's Colab: Free browser-based Jupyter notebook for Python code
  • Vertex AI: Offers a model garden with foundation models
  • Generative AI Studio: For creating and deploying GenAI models
  • Gen AI App Builder: No-code app creation
  • Palm API: Testing and prototyping with Google's language models

Prompt Designing

• Prompt: Short text that guides the model’s output
• Prominence in GenAI: Allows users to generate their own content by providing inputs

Types of Input and Output Models

• Text-to-Text: Translation, Q&A
• Text-to-Image: Generate images based on descriptions
• Text-to-Video: Generate videos from text inputs
• Text-to-3D: Generate 3D objects from text
• Text-to-Task: Perform defined tasks based on text input

Conclusion

• Generative AI: Creates new content from existing content
• Large language models predict new content based on learned patterns
• Tools and applications make generative AI accessible and useful across various domains

Additional Notes

• Transformers: Use encoder and decoder for tasks
• Hallucinations: Nonsensical outputs due to training on insufficient/noisy data

Key Takeaways

• GenAI is a type of AI that creates new data content based on learned patterns from existing data
• Used in various applications from language translation to image generation
• Tools provided by Google Cloud facilitate the development and deployment of GenAI models