Can a machine become an artist? One of the most beautiful demos at this year’s GPU Technology Conference might help make up your mind.
The DCGAN, or deep convolutional generative adversarial networks, demo has a computer take thousands of works of art that share similar features — say seascapes from Victorian artists. It then processes them using deep learning to create new images in the same style. That’s good news if you’re a fan of a long dead painter, but DCGAN is about much more than pretty pictures.
More Than Pretty Pictures
Based on work by Facebook’s Soumith Chintala, DCGAN demonstrates unsupervised learning, considered by many to be the holy grail in the quest for true AI.
To create the DCGAN demo, we trained the deep learning system with thousands of unlabeled images. It analyzed the images and identified features within the data so it was able to understand the difference between, say, a painting of a landscape and a still life while not being aware of these labels.
Then, DCGAN takes things a step further — creating its own original artwork based on that understanding of the different features in the pictures from which it’s learned.
This is a big deal to those following the fast-growing field of deep learning. Deep Learning is a branch of artificial intelligence that lets computers solve problems that are too complex for conventional programming.
The clue to how deep learning works is in the name: systems learn from experience, much like people do. Thanks to its affinity with the parallel architecture of the graphics processing unit, deep learning is massively accelerated by GPUs (see “Accelerating AI with GPUs: A New Computing Model”).
Finding Connections No One Knew to Look For
So, why is unsupervised learning such a big deal? Training any deep learning system involves feeding it massive amounts of data.
Supervised learning uses data human beings have already labeled. Unsupervised learning, by contrast, uses data that no one has labeled.
Let’s say you’re training a deep learning-based system to recognize images of cats. Using supervised learning, you would label the images on which you’re training the system to say which ones are cats and which ones aren’t.
In unsupervised learning, there are no labels. The system has to figure things out for itself by sorting the data. This approach is powerful because it can spot patterns in data without being directed by a developer on what to look for.
It’s not limited by the labels a data scientist has provided, so it can identify connections no human would think to look for. Unsupervised deep learning has proved particularly effective with once insoluble problems.
Changing the World
At GTC, DCGAN is making art. Soon systems based on the same principles will be helping us solve some of our most complex challenges. From computer vision and natural language processing to medical diagnosis, unsupervised deep learning accelerated by GPUs is changing the world.