Computational creativity: artificial intelligence models for the generation and classification of affective artwork and their human evaluation

Abstract

Visual artwork is among the most salient forms of human expression. From prehistoric cave paintings to Renaissance and modern art, paintings have been a powerful medium for expressing emotions. With the advent of computing and artificial intelligence, visual arts may no longer be exclusive to human creativity. Computational creativity involves the study of creative endeavors ranging from creative writing, poetry, painting, music, and science to sports through computational approaches. On the intersection of art and computer science, this thesis involves implementing artificial intelligence models to generate and classify affective artwork and their human evaluation. Rooted in the WikiArt data of over 80,000 paintings and their emotion labels from ArtEmis, we implement Generative Adversarial Networks for generating paintings with desired emotional content. We first experiment with two broad classes of emotions (positive and negative) to further deal with nuanced affective categories, viz. amusement, awe, contentment, excitement, anger, disgust, fear, and sadness. Besides computationally generating affective artwork, we also implement classification models and validate their performance using relevant metrics. Projected GANs, StyleGAN2-ADA, and StyleGAN3 are employed for generating artwork for binary and multi-class models to achieve an FID score of 7.84 for the StyleGAN2-ADA architecture. ResNet50-V2 presents the highest accuracy for the binary classification experiment at 72%. Beyond the computational evaluation of the generated artwork, we created a ‘Turing Test for Artist.’ This test randomly presents images of human artwork, and those made with artificial intelligence to a human evaluator and registers their assessment. For every image, the test also records the binary human assessment of the affective content of the artwork. We assess the quality of the generation and classification models after conducting the Turing Test with a sizeable number of evaluators. We conclude that while the artificial intelligence approach is capable of producing affective artworks that compete with human creativity, it is far from replacing it.