AI is Exploding!
Learn about the advancement in artificial intelligence and how GANs give valuable insights into what we can do with AI.
We'll cover the following
Rise of ML and AI
Machine learning (ML) and artificial intelligence (AI) has exploded in the last few years with amazing achievements being reported in the news every few months.
Not only does your smartphone understand you when you talk to it, but it is also pretty good at translating between many human languages. Self-driving cars are now able to drive as safely as humans, and machines can now diagnose some diseases more accurately and sooner than experienced doctors.
The ancient Chinese game of Go has been played for 3,000 years, and although the rules are simpler than chess, the game itself is much more complex requiring longer-term strategies. Only very recently, researchers were able to guide a machine learning system to play and beat a world champion for the first time. That machine learning system also discovered new game strategies that went undiscovered for Go’s entire 3,000 year history.
Beyond learning to perform a task, the discovery of new strategies is a profound success in the field of machine learning.
Creative AI
In October 2018, the prestigious auction house Christies sold a portrait for $432,500. That painting was not painted by a person, but was generated by a neural network. A portrait painted using AI and sold for almost half a million dollars is a milestone in the history of art.
The neural network used was trained using a new and exciting technique called adversarial training. The architecture is called a generative adversarial network, or GAN.
GANs are attracting a lot of interest, particularly from the creative technology sector, because they can create images that look very plausible. Those images are not made by simply copying and pasting parts of the training examples, and they’re not mushy averages of the training data either. This is what sets GANs apart from most other forms of machine learning. GANs are learning to create images at a level above merely replicating or averaging training data.
Yann LeCun, one of the world’s leading researchers on neural networks called GANs “The coolest idea in deep learning in the last 20 years.”
GANs are new
Compared to the decades of research and refinement behind traditional neural networks, GANs only came to prominence in 2014 with Ian Goodfellow’s now seminal paper Generative Adversarial Networks.
This means GANs are very new, and the creative possibilities are only starting to be explored.
It also means we don’t yet fully understand how to train them as effectively as we can with traditional networks. When they work, they work spectacularly well, but too often they fail. Research into how GANs work, and why they can fail, is currently very active.