Generative Models Flashcards
(12 cards)
What is a generative model?
A model that learns the data distribution p(x) to create novel content such as images, text, audio, or video.
Name three applications of generative models mentioned in the lecture.
Data augmentation, dataset synthesis, and art creation (also code generation, text generation, audio synthesis).
Contrast discriminative and generative models.
Discriminative models learn p(y|x) for prediction tasks; generative models learn p(x) to model data; conditional generative models learn p(x|y).
List the types of generative models covered in the lecture.
Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), and diffusion models.
What principle underlies Variational Autoencoders?
VAEs use an encoder-decoder and maximize the Evidence Lower Bound (ELBO) on data log-likelihood via variational inference.
What limitation of VAEs was highlighted?
VAE-generated samples tend to be blurrier and lower quality compared to those from GANs and diffusion models.
How does a GAN work?
A generator network synthesizes data from noise to fool a discriminator network, which learns to distinguish real from fake data.
What is the training objective for GANs?
A minimax game where the generator tries to maximize the discriminator’s error while the discriminator tries to minimize it.
What is a DC-GAN?
A GAN that uses deep convolutional neural networks for both generator and discriminator, leading to higher-quality image generation.
What defines a conditional GAN?
A GAN where both generator and discriminator receive class labels y as input, enabling conditional data generation p(x|y).
What is a diffusion model?
A generative model formulated as a Markov chain, trained via variational inference to learn reverse denoising steps.
What terms make up the ELBO in diffusion models?
A prior matching term (no trainable parameters) and denoising matching term learned with a decoder.
