Summary
Highlights
Dr. Gwendolyn Stripling introduces the course on Generative AI. The course will define generative AI, explain its workings, describe various model types, and discuss applications.
AI is a branch of computer science focused on creating intelligent agents that can reason, learn, and act autonomously. Machine learning is a subfield of AI where models learn from data to make predictions without explicit programming. Supervised learning uses labeled data for prediction, while unsupervised learning discovers patterns in unlabeled data. Deep learning, a subset of machine learning, uses artificial neural networks inspired by the human brain to process complex patterns, often employing semi-supervised learning with both labeled and unlabeled data.
Machine learning models can be generative or discriminative. Discriminative models classify or predict labels for data points (e.g., identifying an image as a dog). Generative models create new data instances based on learned probability distributions (e.g., generating a picture of a dog). Generative AI produces new content like natural language, images, or audio, while non-generative AI typically outputs numbers or classifications.
Traditional programming involved hard-coding rules, while neural networks learn from examples to make predictions. Generative AI allows users to create new content such as text, images, and audio. Large Language Models (LLMs) like PaLM and LAMBDA ingest vast amounts of internet data to build foundation models, enabling users to generate content simply by asking questions.
Generative AI creates new content by learning from existing content through a training process, resulting in a statistical model. This model then predicts responses when given a prompt, generating new data similar to its training data. LLMs are a type of generative AI that produce novel combinations of text, while generative image models take images as input and can output text, other images, or videos. Generative language models take text as input and can output more text, images, audio, or decisions, using pattern matching systems to predict what comes next.
The power of generative AI, particularly in natural language processing, stems from transformers, which consist of encoders and decoders. Hallucinations in transformers are nonsensical or grammatically incorrect words or phrases generated by the model, often due to insufficient, noisy, or context-lacking training data or lack of constraints. Prompt design, the art of crafting effective inputs, is crucial for controlling model output.
Generative AI includes various model types: Text-to-Text (e.g., translation), Text-to-Image (e.g., creating images from descriptions using methods like diffusion), Text-to-Video and Text-to-3D (generating video or 3D objects from text), and Text-to-Task (performing defined actions based on text input, such as answering questions or navigating UIs).
Foundation models are large AI models pre-trained on massive datasets, adaptable to a wide range of tasks like sentiment analysis or image captioning. They have the potential to revolutionize industries. Vertex AI features a Model Garden with foundation models like PaLM API for language and Stable Diffusion for vision. Examples of Generative AI applications include code generation, where tools like Bard can debug, explain, translate, and generate code, and Generative AI Studio, which helps developers explore, customize, and deploy models. Generative AI App Builder allows code-free creation of AI apps, while PaLM API and Maker Suite enable quick prototyping and experimentation with Google's LLMs and tools for training, deployment, and monitoring.