Creative Art with Deep Learning
How to embrace deep learning to be more creative on making artworks
“Can we create something that is in itself creative?”, this is an ultimate question in the Artificial Intelligence world. The generative revolution seems already started, we are now able to tell machines to paint artwork with any given style and compose music that is pleasant to listen to, even this technology is far from mature but it already created a lot of anxiety in artists.
In this post, I will first explain how the current generative modeling (one kind of deep learning algorithm) works in plain text, and then point out how we could leverage this new tool to lift our work to a new level instead of being afraid of it.
Generative Modeling
Generative modeling is the rising star technology for producing creative work in deep learning realm, the typical examples are:
(1) Generate human face images for non-real people.
(2) Generate images for car, phone, building or any object categories.
(3) Paint original artwork in different styles.
See the picture of a small town get transferred to different painting styles.
(4) Music creation
(5) Story writing
Idea behind scene
The idea of generative modeling is not so much different with other deep learning algorithms, you first need to collect a bunch of sample data, train the model to learn the data distribution and then use the model to produce some output. (see image below)
The generative model could produce different but similar results every time, the model learns the data distribution (probability) of the sample datasets and generates random sampling from the distribution it learned.
How could we leverage Deep learning models when creating artwork?
Once we understand how the generative modeling works, this new technology doesn’t seem intimidating or threatening. As we saw from the previous section, the art work produced by AI models come from the data distribution of the sample artwork, so the limitation is it can only create similar artworks but can’t go beyond.
Deep learning algorithms need follow certain rules when creating things, it’s either limited by the data distribution from the dataset or regulated by the rules you defined (ex, Alpha Go, is built on learning by rules not past human games). Once we understand this, we could leverage it in our favor.
According to above characteristics, I think we could apply deep learning in below two scenarios to advance our artwork:
- Run quick experiments to find inspiration.
When creating artwork, no matter if it’s painting a picture or composing a music, we often try it in different styles (ex: editing folk music to Rock or Hip hop), see how it looks and then decide what’s next. A great strength of generative modeling is applying different styles on existing artwork, which can shorten the experimentation step and help artists to find inspiration quicker.
2. Define Learning Rules Innovatively
For the artists who want to make something that didn’t exist before, generative modeling could also work. Instead of telling the algorithm to learn from a set of existing music, artists could describe what the music looks like in their mind, and let the algorithm explode for them.
Just like AlphaGo could find new winning strategy with given play rules of Go, we could have the model compose a new music Genre by defining a set of features which doesn’t exist before, one example feature sets could be: beats like Rock, more quadruplet and duplet, etc
As we can see, AI (generative model) is not that crazy or scary, once we understand the theory, we could tame it and let it work for us. I’m a firm believer that AI is just another great invention to unchain our imagination, with proper handling, it could realize our wildest dreams!