What Can We Learn About Creativity from Deep Neural Networks

Deep neural networks (DNNs) have sparked a revolution in artificial intelligence (AI), enabling effective natural language processing (NLP), image classification, self-driving vehicles, and many other applications. DNNs, which were originally inspired by a crude model of the operation of nerve cells, have the key property that their behavior is governed much more by the data that is fed to them than by the DNN algorithm itself. Even a good understanding of the algorithm is nearly useless for understanding the behavior of a DNN once it has been trained. Their ability to surprise is remarkable and has led to their use by artists as a novel artistic medium. In this talk, I will show that nondeterminism, feedback, and DNNs combine to yield a plausible model of “machine creativity“ that may lend some insight into human creativity. The talk is based on my recent book, The Coevolution (MIT Press, 2020).