UDS21403J – working with big data

Unit – 13

SRM Institute of Science and Technology

1

Generative Adversarial Networks (GAN)

Generative Adversarial Networks (GANs) are a type of machine learning model used for generative modeling. GANs were introduced in 2014 by Ian Goodfellow and his colleagues at the University of Montreal. GANs are composed of two neural networks, a generator and a discriminator, which are trained in a adversarial setting to produce new data that is similar to the training data.

The generator network is responsible for creating new samples that resemble the training data. It takes random noise as input and produces an output that ideally is indistinguishable from the real data

The discriminator network, on the other hand, is trained to distinguish between the generated samples and the real samples. It takes both the generated and real data as input and produces a binary output indicating whether the input is real or generated. The two networks are trained in an adversarial setting where the generator tries to produce samples that can fool the discriminator into believing they are real, while the discriminator tries to correctly identify which samples are real and which are fake.

​

​

The training process for GANs involves an iterative process where the generator and discriminator networks are trained in alternating rounds. In each round, the generator creates a set of new samples, which are then presented to the discriminator along with an equal number of real samples from the training data. The discriminator is then trained to distinguish between the real and generated samples, while the generator is trained to produce samples that can fool the discriminator. This process continues until the generator produces samples that are indistinguishable from the real data.

GANs have been used in a wide range of applications, including image synthesis, video synthesis, music synthesis, and text synthesis. They have shown remarkable success in generating realistic images, and have been used to create deepfakes, which are synthetic images or videos that appear to be real.

GANs are also used for data augmentation, which involves generating new data to expand the size of the training set, and for style transfer, which involves transferring the style of one image onto another. Despite their success, GANs are still a challenging area of research and require careful tuning and monitoring to ensure that the generated data is of high quality and does not exhibit any bias or unfairness.

​

Generative Adversarial Models Overview, Discriminative vs Generative Modelling, Examples of Generative models

Generative Adversarial Models Overview:

​

Generative Adversarial Networks (GANs) are a type of generative model that uses a two-part network to generate new data that is similar to the training data. The first part of the network, called the generator, takes random noise as input and produces a sample that resembles the training data. The second part of the network, called the discriminator, takes both the generated and real data as input and produces a binary output indicating whether the input is real or generated. The two parts of the network are trained in an adversarial setting, where the generator tries to produce samples that can fool the discriminator into believing they are real, while the discriminator tries to correctly identify which samples are real and which are fake.

Discriminative vs Generative Modelling: Discriminative models are used to predict a target variable given a set of input variables. These models learn the conditional probability distribution of the target variable given the input variables.

​

For example, a logistic regression model can be used to predict whether a customer will purchase a product based on their age, gender, and income.

​

In contrast, generative models are used to model the joint probability distribution of the input and target variables. These models can generate new data that is similar to the training data.

For example, a generative model can be used to generate new images that are similar to the images in the training data

Examples of Generative Models:

​

Variational Autoencoder (VAE): VAEs are a type of generative model that uses a neural network to encode the input data into a low-dimensional space, called the latent space. The latent space is then used to generate new data that is similar to the training data. VAEs have been used for image synthesis, text generation, and drug discovery.

​

Pixel CNN: Pixel CNN is a type of generative model that uses a neural network to predict the probability distribution of each pixel in an image, given the values of the other pixels. Pixel CNN can generate new images that are similar to the images in the training data.

​

Auto-Regressive Models: Auto-regressive models are a type of generative model that models the conditional probability of each element in the output sequence given the previous elements. These models have been used for speech synthesis, language modeling, and music generation. Examples of auto-regressive models include WaveNet, a speech synthesis model, and GPT-3, a language model.

​

​

​

​

Generative Adversarial Networks Overview, The Generator Model, GAN’s and CNN’s

The generator model is an essential part of Generative Adversarial Networks (GANs), which are a type of deep learning model used for generative modeling. The generator model is responsible for creating new samples that resemble the training data. It takes random noise as input and produces an output that is ideally indistinguishable from the real data. The generator model is trained to produce samples that can fool the discriminator into believing they are real.

​

Convolutional Neural Networks (CNNs) are a type of neural network that are commonly used in GANs for both the generator and discriminator models. CNNs are particularly well-suited for image-based tasks because they can capture spatial relationships between pixels. In GANs, the generator model typically consists of a series of layers, each of which applies a set of convolutional filters to the input noise. The output of the final layer is an image that is generated by the network.

The discriminator model in GANs also typically uses CNNs. The discriminator model takes both the generated and real data as input and produces a binary output indicating whether the input is real or generated. The discriminator model is trained to distinguish between the real and generated samples.

GANs with CNNs have been used in a wide range of applications, including image synthesis, video synthesis, and 3D modeling.

CNNs allow the models to capture complex patterns in the data, which can lead to more realistic and detailed output. However, training GANs with CNNs can be challenging because they require a large amount of data and computational resources.

Additionally, GANs are prone to instability during training, which can lead to poor quality output or mode collapse, where the generator produces only a limited set of output samples.

Conditional GAN’s, Why Generative Adversarial Networks, Generative Adversarial Networks Training, Loss Functions

Conditional GAN’s:

​

Conditional GANs are a type of Generative Adversarial Network (GAN) that allows for the generation of samples based on specific conditions or labels. In a standard GAN, the generator produces random samples, but in a conditional GAN, the generator takes a conditional input, such as an image label, as well as random noise, to generate samples. This allows for the generation of samples that correspond to specific classes or conditions. For example, a conditional GAN trained on the MNIST dataset could generate handwritten digits of a specific class, such as "7".

Why Generative Adversarial Networks:

​

Generative Adversarial Networks (GANs) are a powerful tool for generative modeling because they can generate new data that is similar to the training data. GANs are particularly useful for image and video generation,

where it can be difficult to manually design features and patterns. GANs are also useful in situations where data is scarce or difficult to obtain.

For example, GANs have been used to generate synthetic medical images for training machine learning models.

Generative Adversarial Networks Training:

​

Training a GAN involves training both the generator and discriminator models in an adversarial setting. The generator tries to produce samples that can fool the discriminator into believing they are real, while the discriminator tries to correctly identify which samples are real and which are fake. The training process typically involves alternating between updating the generator and discriminator models. The generator is updated by backpropagating the error signal from the discriminator back through the generator. The discriminator is updated by backpropagating the error signal from the binary output of the discriminator back through the discriminator.

Loss Functions:

​

In GANs, the loss function is a critical component of the training process. The generator and discriminator models each have their own loss functions. The generator loss function encourages the generator to produce samples that are similar to the training data, while the discriminator loss function encourages the discriminator to correctly identify real and fake samples. The loss functions are typically optimized using gradient descent algorithms such as Adam. Different loss functions can be used in GANs, including the binary cross-entropy loss, Wasserstein loss, and hinge loss. The choice of loss function can have a significant impact on the performance of the GAN model.