Patents, generative algorithms, and innovation frontiers
Can AI recognize innovation? We operationalize arguments from the philosophy of creativity and theory of innovation literatures to argue that three different types of scientific innovation — combinational, exploratory, and transformative — can be distinguished by generative algorithms in an unsupervised, data-driven manner. We apply a customized variational autoencoder (VAE) on a data set of “computing systems” patents to generate and situate patents in an interpretable vector space, consisting of factors of innovation. To test the value of this representation, we explore correlations with economic measures of patent impact as measured by:
i) market value changes as recorded in Kogan et al. 
ii) forward citations
Combinational innovations have grown at an exponential rate while creative innovations remain steady. Creative innovations have greater technological impact, but not market impact. Our paper demonstrates that generative algorithms can identify different categories of human innovation, and underscores the potential utility of generative AI methods for business applications.