This paper, "Quantifying Creativity in Art Networks," by Elgammal and Saleh, proposes a computational framework for assessing the creativity of artistic products, specifically paintings and sculptures. The core idea is to quantify creativity based on two widely accepted criteria: originality (novelty compared to prior works) and influence (impact on subsequent works). The framework doesn't aim to replace art historians but to provide an objective, computational tool that can process vast amounts of artwork in the age of digital humanities.
Here's a breakdown of the key components and contributions:
- Defining Creativity: The authors adopt a product-centric view of creativity, aligned with historical/objective creativity (H-creativity), focusing on the artwork itself rather than the artist's mental process. This aligns with Kant's definition of artistic genius requiring "originality" and being "exemplary."
- Painting Graph Construction:
- A directed graph is created where each node represents a painting, and edges connect paintings based on their creation date (edges point from earlier to later works).
- Edge weights represent the similarity between paintings, calculated using a similarity function
S(.,.). This function can be based on various visual aspects (a), such as color, subject matter, or composition, allowing for multi-dimensional creativity assessment. The paper focuses primarily on features quantifying subject matter and scene composition (using Classeme and GIST features).
- Only the top K most similar prior paintings are connected to any given painting, preventing bias towards modern works with many predecessors.
- Creativity Propagation and the Creativity Implication Network (CIN):
- The core innovation is how creativity is propagated through the network. High similarity between a subsequent painting and a prior one decreases the subsequent painting's creativity (less original) and increases the prior one's (more influential). Conversely, low similarity increases the subsequent painting's creativity (more original) and decreases the prior one's (less influential).
- A balancing function,
B(w), is introduced. This function compares edge weights to a threshold, m(i), specific to each node. Weights above the threshold are considered "high" (indicating influence), and those below are considered "low" (indicating originality). The balancing function essentially converts weights below m(i) to negative values.
- To handle negative weights, the Creativity Implication Network (CIN) is constructed. Negative edges are reversed, and their values negated. This transforms the graph so that higher edge weights always imply lower creativity for the subsequent node and higher creativity for the prior node. The direction of edges in the CIN no longer represents temporal order, but the direction of creativity influence.
- Computing Creativity Scores:
- A recursive formula (Eq. 1) is defined to calculate creativity scores,
C(p_i), for each painting. This formula aggregates a fraction (α) of the creativity scores of a node's outgoing neighbors in the CIN, weighted by the adjusted edge weights. The constant term (1-α)/N reflects a probability that the similarity between paintings doesn't necessarily means a causal influence.
- This formulation is shown to be related to established network centrality algorithms like weighted variants of eigenvector centrality, Hubbell's centrality, and PageRank. A closed-form solution (Eq. 3) is also provided.
- Originality vs. Influence (β parameter):
- The framework can be modified to emphasize either originality or influence using a parameter, β (Eq. 4 and 5). Edges in the CIN are labeled as "prior" or "subsequent" based on the original temporal relationship. β controls the weighting of scores propagated from prior nodes (originality) versus subsequent nodes (influence).
- Visual likelihood and Temporal Prior: The similarity function used in the experiments is based on a Gaussian probability model. The similarity score S(p_j, p_i) represents that painting p_j is created, given the feature vector from a painting p_i. A Gaussian probability density model is applied.
- A temporal prior can be added to the similarity function. The temporal prior decreases the influence of paintings that are very distant in time, as that similarity can be a coincidence or a style revival.
- Experiments and Results:
- Two datasets are used: Artchive (1710 artworks) and Wikiart (pruned to 62,254 western paintings).
- Qualitative Results: The paper presents plots of creativity scores versus time, highlighting paintings that score unusually high or low. Examples include Edvard Munch's "The Scream," Picasso's "Ladies of Avignon," and Malevich's Suprematist works, all of which are recognized as highly creative and influential. The algorithm also identifies incorrectly dated paintings in both datasets, demonstrating its potential for data cleaning.
- Originality vs. Influence Experiment: Analysis of religious paintings from Wikiart shows how adjusting β shifts the focus. Emphasizing originality biases scores towards modern works, while emphasizing influence favors earlier ones.
- Two-Dimensional Creativity: Analysis of portrait paintings using both Classeme and GIST features creates a 2D creativity space, showing how different visual aspects correlate with different dimensions of creativity.
- Time Machine Experiment: A quantitative validation method is introduced. Paintings are "moved" in time (their dates are artificially changed), and the resulting change in creativity score is measured. The results show that moving creative paintings back in time generally increases their scores, and moving paintings forward decreases them, as expected. This is done across various art movements (e.g., Impressionism, Cubism, Renaissance, Baroque) to test the algorithm's consistency.
- Conclusion: The paper concludes that the framework successfully identifies creative artworks based solely on visual analysis and creation dates, without any explicit knowledge of art history. The time machine experiment provides quantitative evidence supporting the algorithm's validity. The framework is also applicable to other creative domains beyond visual art.
In summary, the paper presents a novel, network-based approach to quantifying creativity in art. It leverages computer vision techniques and a clever graph transformation (the CIN) to propagate creativity scores based on originality and influence. The qualitative and quantitative results demonstrate the framework's potential for providing objective insights into the creativity of artworks and its potential applications in digital humanities.