Generative Models: GANs and VAEs Quiz

Generative Artificial Intelligence - Unit 1

• Generative AI is a type of AI system designed to create new content, such as text, images, audio, video, code, or other data types.
• Generative models learn patterns and features from large datasets of images, and create new images that are similar to, or completely different from, the original training data.
• This process generates content akin to image synthesis.
• Generative AI models have both advantages and disadvantages.
• Generative AI models rely heavily on the dataset they are trained on, needing a large dataset to represent the richness and variety of the target picture domain to ensure accurate results, like generating diverse medical images that reflect a wide range of illnesses, organs, and imaging modalities.
• A generative model dataset contains numerous examples of the subject matter, making it a training dataset. Each data point is called an observation.
• The generative modeling process begins by training the model on vast quantities of data, allowing the model to learn patterns and rules governing that data's appearance.
• In the case of horses, the model would learn general rules that govern the appearance of horses, enabling it to generate new horse images that appear realistic.

Generative Adversarial Networks (GANs)

• GANs are a popular and effective type of generative AI model for image creation, comprised of two neural networks:
  • A generator network creates new images.
  • A discriminator network determines if the images are real or fake.
• GANs are trained in parallel using adversarial training.
• The generator attempts to deceive the discriminator, while the discriminator identifies and classifies generated images as real or fake.
• This cycle of training results in increasingly realistic and difficult-to-distinguish images by the discriminator.

Variational Autoencoders (VAEs)

• VAEs are another type of generative AI model for picture synthesis, consisting of:
  • An encoder that learns a compressed representation of input images (latent space).
  • A decoder that uses the latent space representation to generate new images.
• These models allow generating high-quality images featuring intricate features, textures, or patterns and complex visuals, when combined with adversarial training.
• VAEs possess a probabilistic component which enables creating diverse new image variations starting from a single input image. However, VAEs have difficulty producing very realistic images and take longer to create them, encoding and decoding for each generated image.

Autoregressive Models

• Autoregressive models generate images pixel-by-pixel, starting from a seed image.
• The model predicts the next pixel's value based on preceding pixel values.
• Autoregressive models create high-quality images with intricate details but produce them comparatively slowly since each pixel is generated separately. They are effective in generating high-quality images for picture inpainting and super-resolution.
• Compared to GANs, they might struggle in generating extremely realistic images.

Choosing the Right Dataset

• Generative AI models are highly reliant on the training dataset. It must be large enough to represent the target picture domain accurately and contain diverse variations of data for accurate results.

Generative vs. Discriminative Modeling

• Discriminative models predict the probability of an observation belonging to a specific category.
• Generative models are concerned with the probability an observation may be included in a particular class.
• Discriminative models, if able to correctly classify a Van Gogh painting against others, would still be unable to create a similar painting independently from scratch.
• Generative models are capable of generating sets of pixels that closely reflect examples from the training dataset.

Advances in Machine Learning

• Deep learning models have significantly reduced error rates in image classification tasks.
• The ImageNet Large Scale Visual Recognition Challenge marked a breakthrough in this area, achieved by deep convolutional neural networks.

The Rise of Generative Modeling

• Generative models raise ethical concerns regarding the proliferation of fake content online.
• The capability of generating realistic content presents a challenge in distinguishing true from fake information, impacting trust in public communication channels.

The Generative Modeling Framework

• Generative modeling aims to mimic an unknown data-generating rule in a space of two dimensions.
• The framework is used to understand how generative models aim to achieve their results.

Probabilistic Generative Models

• Probabilistic generative models define four key terms: sample space, density function, parametric modeling, and maximum likelihood estimation.

• A sample space encapsulates all possible values an observation can take.

• A density function maps points within the sample space to probabilities between 0 and 1. The cumulative density equals 1.

• Parametric modeling structures the approach to suitable p(model) selection that best reflects the probability distribution of the data from the dataset. It uses a finite number of parameters to define a family of possible density functions.

• Maximum likelihood estimation is a technique for selecting among potential families of density functions, based on values of parameter sets that can accurately reflect observed data.

Practical Application Example

• A practical example involves generating diverse fashion designs from a Fashion Police dataset.