On the topic of creating art, Spanish surrealist painter Joan Miro once said “the works must be conceived with fire in the soul, but executed with clinical coolness.” No matter how much cool compute they may pack, how can today’s AI models hope to access that essential “fire in the soul” when generating their artworks? In a new paper, researchers from Adobe, Georgia Tech, and Facebook AI Research propose a neuro-symbolic hybrid approach to address the challenge of creativity in generative art.
Generative art refers to the creation of artworks using autonomous processes with no direct human control. There are two general classes of generative art: “neural,” where a deep neural network is trained to generate samples from a data distribution; and “symbolic” or “algorithmic,” where a human artist designs the primary parameters and an autonomous system then works within these constraints to generate samples.
The researchers say the neural and symbolic generative visual art approaches had remained largely distinct until this study explored their intersection.
The team first generated massive amounts of training samples using a symbolic approach, then used these to train a data-hungry Generative Adversarial Network (GAN). During the training, an image starts at 4×4 and is doubled every 37k iterations to generate a final 512×512 image. They also used different batch sizes for different image resolutions during training.
The researchers used Circle Packing as their symbolic generator and Progressive GAN as their model architecture. When fed with different input noise vectors, their model was able to produce a variety of promising neuro-symbolic generation samples.
They team crowdsourced humans from Amazon Mechanical Turk (AMT) to compare artworks on various criteria and to evaluate live interactive generative art tools based on traditional symbolic and the proposed neuro-symbolic approaches. The 50 AMT workers were all US-based, had AMT approval ratings of 95 percent or higher, and had previously completed at least 5000 tasks on AMT.
The AMT Workers rated the neuro-symbolic images as more “surprising” and “unusual” than the symbolic images 68 percent of the time, and judged the neuro-symbolic tool as generating more “creative” art 82 percent of the time.
The researchers say the results suggest neuro-symbolic generative art may be a viable new approach and genre. They plan to explore the use of additional symbolic art formats with the goal of eventually training models to discover entirely new artistic styles.
The paper Neuro-Symbolic Generative Art: A Preliminary Study is on arXiv.
Journalist: Yuan Yuan | Editor: Michael Sarazen
This report offers a look at how the Chinese government and business owners have leveraged artificial intelligence technologies in the battle against COVID-19. It is also available on Amazon Kindle.
Click here to find more reports from us.
We know you don’t want to miss any story. Subscribe to our popular Synced Global AI Weekly to get weekly AI updates.