Generative AI is a class of artificial intelligence algorithms designed to generate something new, such as images, music, speech, or text. It is a rapidly evolving field that has produced some of the most fascinating applications of machine learning. It uses inputs in order to create what it has been prompted to do.
Generative AI use cases have exploded recently, and ranges across a variety of from creating realistic images, synthesizing music and speech, to generating synthetic medical data for research purposes. Deepfakes, which use generative models to swap faces in photos and videos, are a controversial but technologically impressive application of generative AI. Other uses include improving computer graphics in video games and movies, creating realistic avatars for virtual reality, and even designing new molecules for drug discovery.
You will understand how more recent developments in generative AI work. We will cover the basics of Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), Transformers, and Diffusion Models. You'll learn about the probabilistic model framework and how these algorithms generate new data instances.
Generative Adversarial Networks (GANs): GANs are currently the most popular type of generative models. They consist of a generator and a discriminator working in tandem to create realistic new data. You'll learn how GANs are trained and how to overcome some of the training difficulties.
Variational Autoencoders (VAEs): Variational Autoencoders, or VAEs, are a type of generative model that are known for their ability to learn complex distributions of data and generate new, realistic samples. They function by having a blend of both deep learning and probabilistic modeling.
Transformers: Transformer models are foundational to many of the breakthroughs in generative AI over recent years. Transformer models are a type of model architecture that use self-attention mechanisms and have shown excellent performance on a variety of tasks where attention to previous context is of utmost importance (particualrly in language models).
Diffusion Models: Diffusion models are a type of generative model that create new images by simulating a random process in reverse. They start with a random image (usually simple Gaussian noise) and gradually refine it into the desired output. They've been gaining popularity due to their ability to produce high-quality images and their interesting training method, which resembles a physical diffusion process. This is the model architecture that powers Stable Diffusion.
Our curriculum was designed for students with at least a pre-calculus background. You'll definitely learn more about math during these lessons, but it shouldn't be anything overly complicated. We've done a careful job balancing both depth of the material vs difficulty.