ARTIFICIAL INTELLIGENCE RATERS: NEURAL NETWORKS FOR RATING PICTORIAL EXPRESSION

Previous studies on classification of fine art show that features of paintings can be captured and categorized using machine learning approaches. This progress can also benefit art psychology by facilitating data collection on artworks without the need to recruit experts as raters. In this study a machine learning approach is used to predict the ratings of RizbA, a Rating instrument for two-dimensional pictorial works. Based on a pre-trained model, the algorithm was fine-tuned via transfer learning on 886 pictorial works by contemporary professional artists and non-professionals. As quality criterion, artificial intelligence raters (ART) are compared with generic raters (GR) created from the real human expert raters, using error rate and mean squared error (MSE). ART ratings have been found to have the same error range as randomly chosen human ratings. Therefore, they can be seen as equivalent to real human expert raters for almost all items in RizbA. Further training with more data will close the gap to the human raters on all items.