Google Cloud Skills Boost - Introduction to GEN AI

• Supervised Learning (Labeled Data)
• Data provided includes tags
• Learns from past examples to predict future outcomes
• Example: Predict amount of tips gave customer based on order type

• No tags or labels on the data
• The goal is to explore and find patterns or groups in the raw data

• Uses artificial neural networks to learn complex patterns
• Neural networks are inspired by the human brain (made of many interconnected neurons)
• Deep learning models have many layers of neurons
• Can work with supervised, unsupervised, or semi-supervised data
• Semi-supervised = small labeled data + large unlabeled data
• Labeled data teaches core concepts; unlabeled data helps generalize
• Generative AI and Large Language Models (LLMs) are subsets of deep learning

Goal: Create new content (text, images, etc.)

How: Learns joint probability p(x, y) and generates new data based on patterns

Example: Generate a dog image based on understanding what a "dog" looks like

• Goal: Classify or label data
• How: Learns conditional probability p(y | x) to predict labels for new inputs
• Example: Given an image, predict if it's a dog or a cat 

• This approach uses training code and labeled data to build a model. Depending on the task, the model can:
• Make predictions
• Classify inputs
• Cluster similar data points

🐱 Before ML: Rule-Based Programming

• In traditional programming, we had to hardcode rules to define what a cat is:
• Type: animal
• Legs: 4
• Ears: 2
• Fur: yes
• Likes: yarn, catnip
• Dislikes: Fred

• Generative AI models take in training code, labeled data, and even unlabeled data of all kinds (text, images, audio, etc.) to build a powerful foundation model.
• This model can then create brand new content — not just recognize patterns but generate text, code, images, audio, and video.

📸 Neural Networks in Action

• With neural networks, we can now show the system images of cats and dogs and ask:
• “Is this a cat?”
• The model doesn't need hardcoded rules — it learns from data and makes a prediction: cat or not a cat.

Generative AI (GenAI) is a type of Artificial Intelligence that creates new content based on patterns it has learned from existing data.

• The learning process is called training, which produces a statistical model.
• When given a prompt, GenAI uses this model to predict likely responses — generating new content.
• It captures the underlying structure of the data and can produce similar, but original, outputs.

• Takes images as input
• Can output images, text, or videos
• Examples: image captioning, super-resolution, video generation, and image completion

• Takes text as input
• Can generate more text, images, audio, or even make decisions
• It learns language patterns from training data
• Predicts how to complete or respond to a sentence

• Driven by transformers, a deep learning architecture
• Key breakthrough in 2018 revolutionized Natural Language Processing (NLP)
• Transformers enable GenAI to understand and generate contextually relevant content

Generative AI can sometimes produce hallucinations — content that sounds plausible but is incorrect or nonsensical.

Reasons include:

• Insufficient training data
• Poor-quality (noisy or dirty) training data
• Lack of context in the prompt
• Inadequate constraints or guidance in prompt design

A prompt is a short input text given to a Large Language Model (LLM) to guide its output.

• Example: “Summarize this article,” or “Write a poem in the style of Shakespeare.”
• Prompt design is the art of crafting effective prompts to achieve desired outcomes from the model.

Predictive ML Model VS Gen AI Model (Chart)

Predictive ML Model VS Gen AI Model (Chart)

• Takes natural language input and produces text output
• Learns to map one text to another
• Common use cases:

1. Translation (e.g., English ↔ Spanish)
2. Summarization
3. Text generation (e.g., writing assistance)

• Trained on large datasets of images with text descriptions
• Generates realistic or stylized images from a text prompt
• Diffusion models are often used in this process
• Example: "A cat wearing sunglasses riding a skateboard"

• Generates videos from text input
• Input can range from a simple sentence to a detailed script
• Example: "Show a sunrise over the mountains with birds flying by"

• Generates 3D models from a user’s text description
• Used in game design, virtual worlds, or AR/VR applications
• Example: "A futuristic flying car with neon lights"

• Executes specific tasks based on text input
• Can perform actions like:

1. Answering questions
2. Running searches
3. Making predictions
4. Automating workflows (e.g., navigating UIs, editing documents)

• Large-scale models trained on massive datasets
• Adaptable to many downstream tasks (via fine-tuning or prompt engineering)
• Tasks include:

1. Sentiment analysis
2. Image captioning
3. Object recognition
4. Question answering

• Industry Impact:

1. Healthcare, finance, customer service, fraud detection, and more

• Example platform: Vertex AI’s Model Garden with support for both:

1. Language models (chat, code, summarization)
2. Vision models (image captioning, embeddings, diffusion models)

A platform that enables developers to quickly explore, customize, and deploy generative AI models on Google Cloud.

Key Features:

• 📚 Library of pre-trained models
• 🛠️ Tools for fine-tuning
• 🚀 Deployment to production
• 💬 Community forums for collaboration and idea sharing

Great for developers looking to get started with generative AI using ready-to-use resources.

Helps build AI-powered search and conversational agents for both customers and internal teams.

Use Cases:

• 🤖 Chatbots
• 🧠 Digital assistants
• 🔍 Custom search engines
• 📚 Knowledge bases
• 🎓 Training and learning applications

Ideal for enhancing customer experience and internal workflows through intelligent, conversational interfaces.

Gemini goes beyond text and enables multimodal capabilities, including:

• 🖼️ Image analysis
• 🔊 Audio understanding
• 💻 Code interpretation

A versatile tool for tasks requiring visual, audio, and programming intelligence — all from a single model.