Generative adversarial networks and the promise they hold for artificial intelligence

What is it about?

Machine learning (ML) has become an essential tool in many fields. However, one of the main downsides of ML models is that they have to be trained. And this is mostly done by a human, who provides the models with manually labelled data, first. The ML model learns from (or is "trained on") this data. Later, it can classify and process new inputs on its own. But what if ML models could work without any prior "training"? Such models do exist. They are called generative models. Of these, generative adversarial networks (GANs) have become the most successful. GANs are mostly used for image generation and analysis. But they can be trained with little data and achieve impressive results. Also, they do not need human supervision! Thus, their use in other applications holds promise as well. This review provides a general overview of GANs. The authors cover what GANs are, how they work, and the main challenges behind them.

Featured Image

Photo by Markus Winkler on Unsplash

Photo by Markus Winkler on Unsplash

Why is it important?

GANs represent a unique and promising approach to artificial intelligence. Their performance has skyrocketed over the past few years. Today, GANs can produce images which are so realistic that even humans have trouble telling if they are real or fake. In turn, GANs' ability to generate fake data can prove very useful. They can be used to train other ML models in cases where real data is difficult to come by. For example, one can use GANs to produce fake photographs of faces to train models for facial recognition. And since fake photos are being used, this would relieve many privacy concerns related to the usage of personal data. GANs can also be useful for researchers. They can be used to simulate experiments in fields like chemistry and materials science. Researchers can learn from these simulations. Then, they would need to replicate only those experiments that seem promising in real time. This would reduce costs and save time. New discoveries and AI tools can thus be made more quickly! KEY TAKEAWAY: This review article on GANs will acquaint researchers with these promising models and their current limitations.