The classic course on Generative AI by Martin Musiol

Explore how generative AI sits among Gartner's top strategic tech trends, generating novel content from data and accelerating R&D in medicine and product design.

Explore generative models like neural style transfer and deep dream, and learn how NFT minting on blockchain uses block hashes and proof of work or stake.

Explore where generative AI sits in the artificial intelligence landscape, as an overlap of machine learning, deep learning, and non-self-learning components like expert systems.

Trace the evolution of deep generative models from autoencoders to variational autoencoders and GANs, and see how encoders, decoders, and latent spaces enable reconstruction, interpolation, and feature-level generation.

Explore how generative adversarial networks generate data and model complex distributions, learning internal representations for unsupervised learning across images, text, audio, video, and time series.

Explore how a generative adversarial network trains a generator and a discriminator in an adversarial loop, using latent seeds to produce samples and distinguishing real from fake toward Nash equilibrium.

Explore neural networks from computing graphs to multilayer networks, training with forward and backward passes and backpropagation, gradients, and supervised learning, including convolutional networks for image tasks.

Demonstrates a simple GAN in TensorFlow that generates 28 by 28 handwritten digits using a generator and discriminator, trained with dropout and binary cross entropy.

Generate 3d objects with a generative adversarial network and volumetric convolutional networks, exploring latent space, chairs and cars, with pix to vox for single- and multi-view reconstruction.

Brainstorm 3D object generation ideas using fashion and interior design data sets, and explore text and image guided 3D avatar generation with differentiable rendering and losses.

Explore interactive image generation using descriptive labels to guide conditional GANs, producing photorealistic scenes with Nvidia Canvas and related models.

Watch a step-by-step demonstration of a conditional gan, where labels are embedded and fused with a 100-dim latent vector to generate 28×28 handwritten digits, and the discriminator uses image-label pairs.

Train a conditional gan with a fixed discriminator and a trainable generator to generate digit images. Track accuracies and losses, and plot results.

Explore how GauGAN uses semantic image synthesis with a spatially adaptive normalization layer to normalize with mean and standard deviation and learnable alpha and beta.

Use data augmentation to balance cancer image datasets and improve model accuracy, generating diverse malignant and benign samples with GANs, including adaptive discriminator augmentation (ADA), while addressing privacy concerns.

Explore image data augmentation techniques, including geometric transforms, color space changes, random erasing, and image mixing, with extensions to text and audio data and practical tools like deep augment package.

Demonstrate adaptive discriminator augmentation in a generative adversarial network to generate augmented bird images from Caltech birds. Explore accuracy targets, integration steps, and evaluation metrics like inception distance and kid.

Explore data augmentation with generative models, emphasizing dataset similarity, effectiveness in low data regimes, bias considerations, and the future role of synthetic data.

Examine the concerns surrounding generative AI, such as deepfake image synthesis and fake landing pages, and explore detection research, datasets, and public benchmarks.

Explore noteworthy gan architectures that shaped generative ai, including dcgans, conditional gans, stacked gans, and pix2pix image-to-image translation. Discover how Wasserstein gans improve training stability and sample quality.

Demonstrate cycle-consistent image translation with CycleGAN, explore progressive GANs, and review style gun and big gun architectures for high-resolution, high-fidelity image generation.