The Origin and Characteristics of Generative Adversarial Networks: GANs

2019-12-28

Artificial intelligence is capable of generating language similar to that of humans or generating images of people that are difficult to distinguish from real-life photos. Typically, these systems are built on the foundation of Generative Adversarial Networks (GANs), which are artificial intelligence models consisting of a generator that creates samples and a discriminator that attempts to differentiate between generated and real samples. This unique arrangement allows GANs to achieve impressive media synthesis feats, from composing melodies and swapping sheep with giraffes to creating illusions of ice skaters and football players. In fact, it is due to this ability that GANs have been used to create problematic content such as deepfakes, a media tool that places one person in existing media and replaces them with another. Yann LeCun, the Director of Artificial Intelligence Research at Facebook, has referred to the evolution of GANs as one of the most interesting ideas of this century. Despite some challenges and shortcomings, GANs remain one of the most widely used neural network architectures today. The idea of pitting two algorithms against each other originated from Arthur Samuel, a prominent researcher in the field of computer science who is credited with popularizing the term "machine learning." During his time at IBM, he designed a checkers game called "Samuel's Checkers Game," which was one of the earliest successful examples of self-learning games, partially achieved by estimating the probability of winning for each side in specific positions. Later, in 2014, Ian Goodfellow, a research scientist at Google Brain, and his colleagues published a groundbreaking research paper titled "Generative Adversarial Networks," marking the formal birth of generative networks. GANs are trained in an unsupervised manner, meaning they can infer patterns in a dataset without referencing known, labeled, or annotated results. Interestingly, the discriminator's feedback informs the generator - whenever the discriminator correctly identifies a synthetic work, it instructs the generator on how to adjust its output to make it more realistic in the future. In practice, GANs have several drawbacks due to architectural issues. Training the generator and discriminator models simultaneously is inherently unstable. Parameters can sometimes oscillate or become unstable, which is not surprising as the nature of the optimization problem being solved changes with each parameter update. Another scenario is when the generator collapses and starts generating data samples that are visually similar. Additionally, there is a risk of the generator and discriminator overpowering each other. If the generator becomes too precise, it exploits weaknesses in the discriminator to produce poor results, while if the discriminator becomes too precise, it hinders the generator's progress towards convergence.

RECENT AI NEWS

RECENT AI TOOLS