Modeling and Forecasting Art Movements with CGANs

• DOI: 10.48550/arxiv.1906.09230
• Publication Date: 2019-01-01
• Authors: Edoardo Lisi; Mohammad Malekzadeh; Hamed Haddadi; F. Din-Houn Lau; Seth Flaxman
• Keywords: Sequence (biology); Generative model; Data set

Conditional Generative Adversarial Networks~(CGAN) are a recent and popular method for generating samples from probability distribution conditioned on latent information. The information often comes in the form of discrete label small set. We propose novel training CGANs which allows us to condition sequence continuous distributions $f^{(1)}, \ldots, f^{(K)}$. This generate distributions. apply our paintings artistic movements, where each movement is considered be its own distribution. Exploiting temporal aspect data, vector autoregressive (VAR) model fitted means that we learn, used one-step-ahead forecasting, predict future art $f^{(K+1)}$. Realisations this can by CGAN "future" paintings. In experiments, methodology generates accurate predictions evolution art. set consists large dataset past While there no agreement exactly what current period find ourselves in, test plausible candidate sets present art, show mean distance small.