Why Do We Need Generative Adversarial Networks (GANs) To Further The Application Of Machine Learning?

Machine Learning has become one of the influential disruptive technologies of the 21st century. With applications extending from precise diagnosis of skin diseases, detection faults in credit lending systems to recommendations on streaming channels and gaming, this technology is omnipresent. However, there is are the darker, malignant side of it too. One of the most concerning challenges it fooling the existing algorithms of neural networks by adding a small amount of noise into the original data. And after many iteration cycles and feedbacks, the same model can now produce counterfeit data (Deepfake) or wrong output. This occurs because, after adding noise, the model has higher confidence in the wrong prediction than when it is predicted correctly. Hence, to tackle such Deepfakes and misclassification of data, researchers are leveraging on the neural technology, which helped to create dupes and errors in the first place: Generative Adversarial Networks (GANs).

What are GANs?

The concept was first elucidated by Ian J. Goodfellow and co-authors in the article Generative Adversarial Nets in 2014. Yann LeCun, Facebook's Director of AI Research, mentions GANs as 'the most interesting idea in the last ten years in machine learning.' This powerful class of neural networks that are used for unsupervised learning takes up a game-theoretic approach, unlike a conventional neural network. The main goal of GANs is to learn from a set of training data and generate new data with the same characteristics as the training data.  It is composed of two neural network models, a generator and a discriminator. These two adversaries compete with each other and can analyze, capture, and copy the variations within a dataset. In general parlance, the Generator is trained to produce believable fake data from a random source, while the Discriminator is trained to distinguish the Generator's simulated data from real data. The Generator is trained while the Discriminator is idle and vice-versa. In some cases, GANs also consist of multiple discriminators and generators, like MD-GAN (Multiple Discriminator Generative Adversarial Network), SGAN (Semi-Supervised Adversarial Network), and many more.

Application

The most familiar adopter of this technology are industries relying on computer vision. Some of the common examples of applications of GANs are:

• Generation of synthetic training data for machine learning models in case training data is insufficient or collecting it is too costly.

• Generation of human faces, objects in 2D and 3D, realistic photographs, anime characters, and music.

• Identifying instances of fraud when an adversarial attack is launched by hackers to seek information.

• Detection of tumor in human bodies by comparing images with a dataset of images of healthy organs.

• Assisting in the drug discovery process by generating molecular structures for medicines using an existing database to find new compounds that can potentially be used to treat new diseases.

• Media Translations like image-to-image translations, semantic image-to-photo translations, and text-to-image translations.

• Editing photographs by denoising images and enhancing the existing image data using super-resolution, photo blending.

• Identifying criminals that might have undergone surgeries to modify their appearance, using image reconstruction, Face Frontal View Generation.

• Predicting the next frame in images and videos.

• Converting Human photographs into Emojis, Bitmojis (like in Apple iPhone, Snapchat), or applying filters of Instagram, Faceapp, and so on.

• Creating Image style transfers or audio style transfers.

Outlook:

While the hype around GANs may seem to exaggerate, this field promises of numerous applications and yet untapped potential. With further research, training, and development, it is poised to create multiple advantages for several industries.