Let us commence by defining generative adversarial networks.
Generative Adversarial Networks (GANs) use a pair of neural networks that oppose each other to produce novel, synthetic data that can be difficult to differentiate from authentic data. The networks compete with each other, which is why they are referred to as ‘adversarial’. GANs are highly effective in producing dynamic images, sound, and written content.
Generative Adversarial Networks (GANs) have enormous potential to reproduce any data distribution, making them useful for both ethical and malicious intentions. GANs can learn to produce synthetic environments that are uncannily akin to actual reality in any form, including visual, aural, and textual.
Generative Adversarial Networks (GANs) are essentially an AI-based art form capable of generating high-quality outputs, but they can also be misused to create deep fakes or other types of media counterfeits.
The ‘Magic’ of GANs
The notion of adversarial training is highly intriguing to me. Similar to backpropagation, which is a basic yet innovative approach that has enhanced the efficacy and ubiquity of neural networks, adversarial training also has a certain appeal in its simplicity and is an essential milestone in the realm of machine learning, especially for generative models.
At times, the capacity of Generative Adversarial Networks to generate novel content may appear “magical”.
In the next passages, we will investigate the mathematical and conceptual framework for these projections, as well as the logic behind them. Furthermore, we will examine the key principles and models that underpin these projections to reveal the truth behind the hype. As we proceed, we will build and clarify the justification that serves as the basis for these concepts.
The Role of GANs:
The elements of a Generative Adversarial Network comprise:
Generative:
A generative model defines the process of data generation.Adversarial:
The model is refined using an adversarial training approach.- Networks: The training process employs Artificial Intelligence (AI) systems called deep neural networks.
Every Generative Adversarial Network (GAN) comprises of a generator and a discriminator. The generator produces artificial data samples, like images or audio, and attempts to trick the discriminator. Meanwhile, the discriminator’s role is to differentiate between genuine and synthetic samples.
In the training phase, a generative and discriminative neural network model engage in an antagonistic competition. With successive iterations between the two models, they enhance their individual abilities and become more adept at their respective roles.
The generator’s task is to produce fresh data examples which the discriminator then scrutinizes for authenticity. In other words, the discriminator validates that the newly generated data instances are credible representatives of the training dataset.
Our goal is to reproduce the handwritten numerals found in the MNIST dataset using only genuine data. The discriminator’s role is to differentiate between legitimate images in the MNIST dataset and the artificial ones. While the discriminator evaluates these images, the generator generates fresh synthetic images with the intention that humans would struggle to distinguish between real and synthetic images.
The generator has been developed to generate lifelike handwritten numerals, essentially allowing users to deceive others without being detected. The discriminator is constructed to detect the counterfeit images produced by the generator.
What is the process involved in GANs?
- The generator produces an image based on a sequence of random integers.
- The image, along with a continuous stream of images from the genuine ground truth dataset, is fed to the discriminator.
- The discriminator takes both the genuine and synthetic images and offers probabilities between 0 to 1, where a score of 1 signifies authenticity, for each image.
This results in a feedback loop in which both directions are operational. As follows:
Just as the discriminator is connected back to the genuine ground truth image, similarly the generator is also connected back to the discriminator.
A Generative Adversarial Network (GAN) could be likened to a game of cat and mouse between a forger and a detective. The forger is actively gaining knowledge on how to produce convincing counterfeit notes, while the detective is equally invested in learning how to detect them. This process is ongoing, with both parties consistently improving their tactics. For instance, the detective may still be figuring out the intricacies of counterfeit identification (to continue the analogy, one might say that the government is taking measures against currency that has evaded detection).
The Multilayer Perceptron (MNIST) discriminator network is a typical convolutional neural network that classifies input images into either “real” or “fake” categories. In contrast, the generator’s function can be seen as the opposite of a convolutional network because it takes a vector of random noise and upscales it to generate an image. In contrast, a conventional convolutional classifier downsizes an image to provide a probability, often utilizing techniques such as max-pooling, which ultimately results in a loss of information. On the other hand, the generator generates new information utilizing the random noise vector as an input.
In a zero-sum game, each network has a goal that must be maximised to achieve success. An actor-critic model comprises of two components: the actor, responsible for making decisions based on the environment, and the critic, tasked with evaluating the actor’s performance. The critic’s actions can significantly influence the generator’s performance, and the generator will react accordingly, often leading to a cascade of losses that may be accumulated over time.
