Creativity in the Age of Artificial Intelligence

How concepts like Transfer Learning and Domain Adaptation will drive creativity.

I was reading "Originals: How Non-Conformists Move the World" by Adam Grant in 2018 when I came across an interesting fact on creativity about Galileo and his telescope. Galileo looked through his telescope and claimed that the moon had craters and mountains just the way the Earth did. It was a pretty unconventional claim at the time. The Moon was believed to be a perfectly spherical object as described by Aristotle. To claim that it had an uneven surface was considered blasphemous!

Galileo wasn't the first person to look at the moon through a telescope. There were others before him. They saw the same images of the moon that he did. What differentiated Galileo from the others was his knowledge of another domain (art) and his application of that knowledge to his newly founded discovery.

Chiaroscuro is a technique of drawing 3D images on to a 2D surface by understanding the interplay of light and shadows. Galileo learned it as a child in his art classes. When he looked at the moon via his telescope he noticed patterns similar to the ones he learned in his Chiaruscuro classes. For several days he kept looking at the moon to confirm his hypothesis that the patterns were changing as the source of light (the Sun) changed. With that knowledge, he was able to calculate the height of the mountains and the depth of the craters that helped describe the surface of the moon. It is surprising how the others who were looking at the same images never saw it coming. Galileo's ability to look at the moon differently has helped science grow immensely.

There are lots of scientific discoveries where such creative solutions were found by applying knowledge from one domain to another. Another classic example is the application of Origami, the ancient Japanese art of paper folding, in the understanding of whether a protein can or cannot exist in a 3D structure - protein folding problem.

While this sounds fascinating, what excites me more is that there are techniques in AI called Transfer Learning and Domain Adaptation that do something similar. You learn to solve a problem based on a data set and then fine-tune the learning to adjust to a whole new dataset. It is very much similar to applying patterns from one domain to another.

For instance, you train a model to learn various features about cats. The model learns about a cat's skin color, the shape of its eyes, ears, tail, its front view, side view, etc. You take the learnings from this model and fine-tune the model to learn features of wild cats like lions and cheetahs. Researches have been able to classify the animals correctly with very high accuracy.

This is just a small example of Transfer Learning. What I am really excited about is to see how people from diverse backgrounds start using these techniques to produce creative results that drive scientific discoveries and create new business opportunities for the future. In fact, Andrew Ng, chief scientist at Baidu and professor at Stanford, said during his widely popular NIPS 2016 tutorial that transfer learning will be the next driver of ML commercial success. I can’t agree more.