Exploring the Interaction between Artificial Intelligence and Artistic Creation: a review

Abstract

In an increasingly technological world, Machine Learning (ML) algorithms are revolutionising art, offering new opportunities and challenges, changing, and shaping the way it is created and appreciated. The growing number of examples of collaborative projects between artists and Artificial Intelligence (AI) systems is transforming traditional notions of artistic authorship and intellectual property. It is also raising several concerns and controversies, including the ethics of data use, cultural appropriation, and originality. Its impact also extends to the art market, to the appreciation of the works of art and to the experience of the target audiences who are confronted with the artistic creations resulting from this process.

Artists use a variety of techniques, including supervised, unsupervised and reinforcement learning, to create new and innovative art forms. In supervised learning, the model learns to associate inputs with outputs by training on a dataset of examples. It can be used for the generation of art which is like a particular style or artist. For example, to generate new paintings that are like Picasso's style using a data set of Picasso's paintings. Generative Adversarial Networks (GAN) have been used to create paintings, fashion design, artistic images, and music production. Recurrent Neural Networks (RNNs) have been used in the generation of music, poetry, and text. The Transformer neural network architecture (used in the GPT model) is used for text generation. Convolutional Neural Networks (CNNs) are effective in the extraction of visual features and are therefore used in the generation of visual art, such as paintings and artistic images. In unsupervised ML, an artist can generate new, unique, unpredictable, and never-before-seen images by training a model on a dataset of random images. It applies clustering techniques (using algorithms such as K-means, Hierarchical Clustering, or DBSCAN) and dimensionality reduction (using algorithms such as PCA - Principal Component Analysis, t-SNE - t-Distributed Stochastic Neighbour Embedding, Locally Linear Embedding - LLE) to large image datasets. This can be used to identify patterns and styles without the need for labels or human supervision. In reinforcement learning, the model (agent) learns to take actions based on the feedback it receives in the form of rewards or penalties. For example, the model can be used to evaluate a user's picture, drawing or painting and provide the user with information on how to improve their art based on subjective criteria they have learned and desire. In another example, the agent can be taught to take actions to create paintings, music, or other forms of artistic expression that are deemed to be more appealing based on the rewards it has received. Q-Learning, Proximal Policy Optimisation (PPO) and RL from Demonstrations are some of the algorithms used.