DELAUNAY: a dataset of abstract art for psychophysical and machine learning research
This provides a new dataset for psychophysical and machine learning research, enabling more controlled comparisons between human and artificial learning, though it is incremental as it builds on existing dataset creation efforts.
The authors introduced DELAUNAY, a dataset of abstract art labeled by artist names, to bridge the gap between natural images and artificial patterns for comparing human and machine learning, and they trained a convolutional neural network on it to demonstrate its features.
Image datasets are commonly used in psychophysical experiments and in machine learning research. Most publicly available datasets are comprised of images of realistic and natural objects. However, while typical machine learning models lack any domain specific knowledge about natural objects, humans can leverage prior experience for such data, making comparisons between artificial and natural learning challenging. Here, we introduce DELAUNAY, a dataset of abstract paintings and non-figurative art objects labelled by the artists' names. This dataset provides a middle ground between natural images and artificial patterns and can thus be used in a variety of contexts, for example to investigate the sample efficiency of humans and artificial neural networks. Finally, we train an off-the-shelf convolutional neural network on DELAUNAY, highlighting several of its intriguing features.