How Generative Adversarial Networks Work to Simulate Facial Aging
Deep learning has made tremendous strides in recent years due to?Generative Adversarial Networks?(GANs). GANs dramatically improve the ability of AI to generate realistic images of objects, people, and places. They're also adept at performing image-to-image translation tasks, particularly?age progression imagery.??
Research papers that explore how GANs work can be highly technical. As a result, I have summarized how the technology works below.??
GAN Basics
GANs rely on both generative and discriminative models to produce a specific output. The generative model's acting component is known as the generator. Likewise, the discriminative model's acting component is known as the discriminator. Both components operate as?convolutional neural networks.
The generative model seeks to generate output data based on patterns automatically derived from the input data (the original dataset). The discriminative model aims to classify images as either natural or fake and operates as if they were involved in a zero-sum game.
GANs train both approaches to compete against each other until the viewer cannot tell the difference (hence the name, "adversarial network').
Supervised vs. Unsupervised Learning
The generative model conducts unsupervised learning, independently identifying image features. This model generates new data instances after reviewing the distributive properties of the data.
In contrast, the discriminative model is involved in classifying images and thus performs a supervised learning task. This model generates new data instances by learning about the statistical properties of the original data (not its distributive properties).?
Generator vs. Discriminator Functions
In a GAN, the image generator constantly challenges the discriminator, trying to fool it with images that closely resemble the original.??
The discriminator selects whether the received image is authentic by outputting a value between 0 and 1 (fake or real, respectively). It does so by determining the distance between the distribution of the data generated and the distribution of the actual data (the loss function).??
领英推荐
If the discriminator has determined that the generator has provided it with an implausible image, it penalizes the generator. Once it grades the image, the discriminator reenters its training loop.?
Not incidentally, the image generator also has a training loop. It's constantly updated because it incorporates the discriminator as a constant. This form of indirect feedback improves training effectiveness.
Facial Aging GANs
The ability to simulate facial aging generally rests on Conditional Generative Adversarial Networks (cGANs).
Whereas a traditional GAN offers little control over what images it generates, a cGAN guides the generator on what to do. Indeed, a cGAN operates as a supervised learning model, deriving a set of specific characteristics from an image. In a cGAN, the discriminator does not merely determine whether the image is real or fake; it also determines whether the generated (fake) image data corresponds to a set of features.??
Critics contend that cGAN cannot accurately attend to image quality, aging accuracy, and identity preservation simultaneously. They also assert that a single network cannot learn all the various aging effects between two different age groups.?
Several data scientists have proposed a new?Progressive Face Aging framework?based on a GAN (PFA-GAN). By relying on several sub-networks instead of one, each sub-network can specialize in learning only a few specific aging effects between adjacent age groups.??
This framework also factors age distribution by minimizing age estimation loss during facial generation. It even employs a?Pearson correlation coefficient?to measure aging smoothness between images.
Not surprisingly, the data scientists who produced this framework assert that it already demonstrates superior performance over cGAN-based methods.??
Concluding Thoughts
Although age progression imagery is still relatively new, expect it to improve dramatically in the next few years. As noted above, AI researchers are continually tweaking and refining GAN technology to achieve realistic results. Age-progression technology is already impacting the cosmetic, entertainment, and security industries. Expect the full spectrum of GANs to grow.