A neural network account to Kant's philosophical aesthetics

Peter beim Graben

2024-04-13

Abstract:According to Kant's (1724 -- 1804) philosophical aesthetics, laid down in his Critique of the Power of Judgement (1790), beauty is "subjective purposefulness", reflected by the "harmony of the cognitive faculties", which are "understanding" and "imagination". On the one hand, understanding refers to the mental capability to find regularities in sensory manifolds, while imagination refers to intuition, fantasy, and creativity of the mind, on the other hand. Inspired by the reinforcement learning theory of Schmidhuber, I present a neural network analogy for the harmony of the faculties in terms of generative adversarial networks (GAN) - also often employed for artificial music composition -- by identifying the generator module with the faculty of imagination and the discriminator module with the faculty of understanding. According to the GAN algorithm, both modules are engaged in an adversarial game, thereby optimizing a particular objective function. In my reconstruction, the convergence of the GAN algorithm during the reception of art, either music or fine, entails the harmony of the faculties and thereby a neural network analogue of subjective purposefulness, i.e., beauty.

Neurons and Cognition

What problem does this paper attempt to address?

This paper discusses the application of artificial intelligence (AI) in artistic creation, particularly in simulating Kant's aesthetic theory through Generative Adversarial Networks (GANs). The author points out that according to Kant, beauty is subjective purposiveness, manifested in the harmony of cognitive abilities such as understanding and imagination. In the paper, the author associates the generator module of GANs with imagination, and the discriminator module with understanding, simulating this harmony through the adversarial game between these two modules. The author discusses the application of AI in artistic creation, such as DeepArt's image style transfer and GAN-generated artworks, using music and painting as examples.

The key issue raised in the paper is whether AI-generated artworks can be considered as true works of art. The author cites Helsing's ITPRA theory, which describes the five physiological systems involved in music appreciation, as well as recent experimental results that support music appreciation as a dynamic process involving expectation, tension, response, and evaluation. The paper also explores Mayer and Hanslick's aesthetic viewpoints on music, as well as Michell's foundational role in Kant's aesthetic philosophy.

Using model theory and GAN theory, the author attempts to reconstruct the aesthetic reception concept in Kant's aesthetics, linking the training process of GANs with Kant's idea of the "free play of cognitive abilities." The author argues that this neural network simulation can explain the dynamic interaction between imagination and understanding in the process of art reception, and the resulting aesthetic experience.

In conclusion, the paper aims to address how to explain Kant's aesthetic theory using modern scientific methods, and how AI can embody this theory in artistic creation.