DeepCreativity: Measuring Creativity with Deep Learning Techniques

TL;DR

DeepCreativity tackles automatic evaluation of machine creativity by decomposing creativity into value, novelty, and surprise, each computed with deep learning components and combined into a normalized score $DC(a, TCC) = \alpha_1 V(a, D_v) + \alpha_2 N(a, D_n) + \alpha_3 S(a, G_s)$ where $\alpha_i \in [0,1]$ and $\sum \alpha_i = 1$. Value is captured via a GAN discriminator $D_v$, novelty via a style classifier $D_n$, and surprise via gradient-based changes in a sequential generator $G_s$, all trained on a temporally and culturally defined context (TCC). The paper validates the approach with a case study on 19th century American poetry, showing that novelty and surprise increase with historical distance from the training data and that central movements can balance novelty with value, reflecting creative trajectory over time. This creates a context-dependent, automatic framework for comparing creative artifacts without human judgments and has potential for broad domain applicability, albeit with domain-specific limitations noted for novelty and sequential requirements for surprise.

Abstract

Measuring machine creativity is one of the most fascinating challenges in Artificial Intelligence. This paper explores the possibility of using generative learning techniques for automatic assessment of creativity. The proposed solution does not involve human judgement, it is modular and of general applicability. We introduce a new measure, namely DeepCreativity, based on Margaret Boden's definition of creativity as composed by value, novelty and surprise. We evaluate our methodology (and related measure) considering a case study, i.e., the generation of 19th century American poetry, showing its effectiveness and expressiveness.

Figures (2)

• Figure 1: The average of value, novelty, surprise and creativity computed on a sample from the training set and on 20th century American poems.
• Figure 2: The average of value, novelty, surprise and creativity computed on a sample from the training set and on both 18th and 20th century American poems.