Introduction to Generative AI
Course Overview
- Instructor: Roger Martinez, Developer Relations Engineer at Google Cloud
- Course Focus: Define, explain, describe, and explore applications of generative AI
What is Generative AI?
- Definition: A type of AI that produces various types of content (text, imagery, audio, synthetic data)
Context: AI and Machine Learning
- AI: Branch of computer science creating intelligent agents (systems that reason, learn, and act autonomously)
- Machine Learning (ML): Subfield of AI; programs/systems that train a model from input data to make predictions
Types of Machine Learning Models
- Supervised ML Models: Trained with labeled data.
- Unsupervised ML Models: Trained with unlabeled data.
Examples
- Supervised ML: Learning to predict tips based on bill amount and order type in a restaurant
- Unsupervised ML: Clustering employees based on tenure and income
Deep Learning
- Subset of ML using artificial neural networks (inspired by the human brain)
- Processes complex patterns, can use labeled/unlabeled data (semi-supervised learning)
Generative AI
- Subset of deep learning using artificial neural networks
- Can process both labeled and unlabeled data
- Uses supervised, unsupervised, and semi-supervised methods
- Large Language Models (LLMs): Type of deep learning model within generative AI
Model Types: Generative vs. Discriminative
- Discriminative Models: Classify/predict labels for data points
- Generative Models: Generate new data instances based on learned patterns
Examples of Model Use
- Discriminative Model: Classifies if an animal is a dog or a cat
- Generative Model: Generates a new image of a dog
Generative AI Applications
- Natural Language (text): Speech, text generation
- Images: Image generation
- Audio/Video: Audio synthesis, video generation
Mathematical View
- Not Generative: If output Y is a number (e.g., predicted sales)
- Generative: If output Y is natural language or image (e.g., definition of sales, image generation)
Traditional vs. Generative AI Process
- Traditional ML: Uses training code and labeled data to build a model
- Generative AI: Uses training code, labeled, and unlabeled data to build a foundation model capable of generating various types of content
Generative AI Model Types
- Text to Text: Translates text input to text output (e.g., language translation)
- Text to Image/Video/3D: Generates images, videos, 3D models from text descriptions
- Text to Task: Performs defined actions based on text input
Foundation Models
- Definition: Large AI models pre-trained on vast data, adapted for various tasks (e.g., sentiment analysis, image captioning)
- Vertex AI Model Garden: Provides foundation models like PaLM API for text, and stable diffusion for images
Coding with Generative AI
- Code Generation: Debugging, translating code, generating documentation
- Google's Tools:
- Vertex AI Studio: Customize generative AI models
- Vertex AI Search & Conversation: Build chatbots, custom search engines, etc.
- PaLM API: Access large language models for prototyping
Challenges: Transformer Issues
- Transformer Models: Consist of encoder and decoder
- Hallucinations: Problematic nonsensical phrases generated by the model
Prompt Design
- Definition: Creating a prompt to control the output of a large language model (LLM)
- Importance: Ensures desired output based on input text
Tools and Resources
- Vertex AI Studio: Tools for exploring and customizing models
- Vertex AI Search & Conversation: Building applications with minimal coding
- PaLM API: Prototyping with large language models
- Gemini: Multimodal AI model understanding text, images, audio, and code
Summary
- Generative AI: Creates new content based on learned patterns
- Applications: Wide range of applications from text generation to video production
Thank you for attending the course. Check out our other videos to learn more!