Understanding Aesthetic Evaluation using Deep Learning

Jon McCormack,Andy Lomas

DOI: https://doi.org/10.1007/978-3-030-43859-3_9

2020-04-15

Abstract:A bottleneck in any evolutionary art system is aesthetic evaluation. Many different methods have been proposed to automate the evaluation of aesthetics, including measures of symmetry, coherence, complexity, contrast and grouping. The interactive genetic algorithm (IGA) relies on human-in-the-loop, subjective evaluation of aesthetics, but limits possibilities for large search due to user fatigue and small population sizes. In this paper we look at how recent advances in deep learning can assist in automating personal aesthetic judgement. Using a leading artist's computer art dataset, we use dimensionality reduction methods to visualise both genotype and phenotype space in order to support the exploration of new territory in any generative system. Convolutional Neural Networks trained on the user's prior aesthetic evaluations are used to suggest new possibilities similar or between known high quality genotype-phenotype mappings.

Neural and Evolutionary Computing,Machine Learning

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to automate personal aesthetic judgment in evolutionary art systems. Specifically, the paper explores how to use deep - learning techniques, especially Convolutional Neural Networks (CNN), to assist digital artists in searching in the large parameter spaces of modern generative art systems. Traditional methods such as Interactive Genetic Algorithm (IGA) rely on human subjective evaluation, which limits the possibility of large - scale search because of user fatigue and small population size. The paper proposes using machine - learning methods, especially Convolutional Neural Networks, to learn the artist's personal aesthetic preferences and use this knowledge to help the artist find phenotypes that are more in line with their aesthetic concepts or explore new areas in the design space.

The paper attempts to solve the above problems through the following points:
1. **Dataset**: An evolutionary art dataset created by the well - known artist Andy Lomas was used. This dataset contains 1,774 images, and each image is generated from a genotype determined by 12 numerical parameters.
2. **Dimensionality Reduction**: Algorithms such as t - SNE were used to visualize the genotype and phenotype spaces in order to understand the structures and clustering situations in these spaces.
3. **Classification and Ranking Prediction**: The ResNet - 50 Convolutional Neural Network was used to classify phenotypes and predict aesthetic rankings, achieving relatively high accuracy.
4. **Genotype Space Prediction**: The Tabular model in the Fast.ai library was used to predict phenotype categories and aesthetic rankings based on genotype parameters. Although the accuracy is slightly lower, it still has potential practicality.
5. **User Interface Improvement**: New functions were added to the Species Explorer software. A neural network predictor was used to generate 2D cross - sectional diagrams of the genotype space, showing the predicted categories or rankings at different coordinate positions, helping artists discover possible category transition points.

Through these methods, the paper aims to improve the efficiency and creativity of artists when exploring generative art systems, reduce user fatigue, and help them discover new creative areas.