Quantifying Creativity in Art Networks
Ahmed Elgammal and Babak Saleh
Department of Computer Science
Rutgers University
{elgammal,babaks}@cs.rutgers.edu
digihumanlab.rutgers.edu
Abstract
This paper proposes a computational framework for assessing
the creativity of products, such as paintings,
sculptures, poetry, etc. The proposed computational
framework is based on constructing a network between
creative products and using this network to infer about
the originality and influence of its nodes. Through a
series of transformations, we construct a Creativity Implication
Network. We show that inference about creativity
in this network reduces to a variant of network
centrality problems which can be solved efficiently. We
apply the proposed framework to the task of quantifying
creativity of paintings (and sculptures). We experimented
on two datasets with over 62K paintings to illustrate
the behavior of the proposed framework.
Introduction
The field of computational creativity is focused on giving
the machine the ability to generate human-level “creative”
products such as computer generated poetry, stories, jokes,
music, art, etc., as well as creative problem solving. An
important characteristic of a creative agent is its ability to
assess its creativity as well as judge other agents’ creativity.
In this paper we focus on developing a computational
framework for assessing the creativity of products, such as
painting, sculpture, etc. We use the most common definition
of creativity, which emphasizes the originality of the product
and its influential value (Paul and Kaufman 2014a). In the
next section we justify the use of this definition in contrast to
other definitions. The proposed computational framework is
based on constructing a network between products and using
it to infer about the originality and influence of its nodes.
Through a series of transformations, we show that the problem
can reduce to a variant of network centrality problems,
which can be solved efficiently.
We apply the proposed framework to the task of quantifying
creativity of paintings (and sculptures). The reader
might question the feasibility, limitation, and usefulness of
performing such task by a machine. Artists, art historians
and critics use different concepts to describe pantings. In
particular, elements of arts such as space, texture, form,
shape, color, tone and line. Artists also use principles of
art including movement, unity, harmony, variety, balance,
contrast, proportion, and pattern; besides brush strokes, subject
matter, and other descriptive concepts (Fichner-Rathus
2008). We collectively call these concepts artistic concepts.
These artistic concepts can, more or less, be quantified by today’s
computer vision technology. With the rapid progress
in computer vision, more advanced techniques are introduced,
which can be used to measure similarity between
paintings with respect to a given artistic concept. Whether
the state of the art is already sufficient to measure similarity
in meaningful ways, or whether this will happen in the near
or far future, the goal of this paper is to design a framework
that can use such similarity measures to quantify our chosen
definition of creativity in an objective way. Hence, the
proposed framework would provide a ready-to-use approach
that can utilize any future advances in computer vision that
might provide better ways for visual quantification of digitized
paintings. In fact, we applied the proposed framework
using state-of-the-art computer vision techniques and
achieved very reasonable automatic quantification of creativity
on two large datasets of paintings.
One of the fundamental issues with the problem of quantifying
creativity of art is how to validate any results that
the algorithm can obtain. Even if art historians would agree
on a list of highly original and influential paintings that can
be used for validation, any algorithm that aims at assigning
creativity scores will encounter three major limitations: I)
Closed-world limitation: The algorithm is only limited to
the set of paintings it analyzed. It is a closed world for the
algorithm where this set is every thing it has seen about art
history. The number of images of paintings available in the
public domain is just a small fraction of what are in museums
and private collections. II) Artistic concept quantification
limitation: the algorithm is limited by what it sees, in
terms of the ability of the underlying computer vision methods
to encode the important elements and principles of art
that relates to judging creativity. III) Parameter setting: the
results will depend on the setting of the parameters, where
each setting would mean a different way to assign creativity
scores with different interpretation and different criteria.
However, these limitations should not stop us from developing
and testing algorithms to quantify creativity. The first
two limitations are bound to disappear in the future, with
more and more paintings being digitized, as well as with the
continuing advances in computer vision and machine learning.
The third limitation should be thought of as an advanProceedings
of the Sixth International Conference on Computational Creativity June 2015 39
tage, since the different settings mean a rich ability of the
algorithm to assign creativity scores based on different criteria.
For the purpose of validation, we propose a methodology
for validating the results of the algorithm through what
we denote as “time machine experiments”, which provides
evidence of the correctness of the algorithm.
Having discussed the feasibility and limitations, let us
discuss the value of using any computational framework
to assess creativity in art. For a detailed discussion about
the implications of using computational methods in the domain
of aesthetic-judgment-related tasks, we refer the reader
to (Spratt and Elgammal 2014). Our goal is not to replace art
historians’ or artists’ role in judging creativity of art products.
Providing a computational tool that can process millions
of artworks to provide objective similarity measures
and assessments of creativity, given certain visual criteria
can be useful in the age of digital humanities. From a computational
creativity point of view, evaluating the framework
on digitized art data provides an excellent way to optimize
and validate the framework, since art history provides us
with suggestions about what is considered creative and what
might be less creative. In this work we did not use any such
hints in achieving the creativity scores, since the whole process
is unsupervised, i.e., the approach does not use any
creativity, genre, or style labels. However we can use evidence
from art history to judge whether the results make
sense or not. Validating the framework on digitized art data
makes it possible to be used on other products where no such
knowledge is available, for example to validate computergenerated
creative products.
On the Notion of Creativity
There is a historically long and ongoing debate on how to
define creativity. In this section we give a brief description
of some of these definitions that directly relate to the notion
we will use in the proposed computational framework.
Therefore, this section is by no means intended to serve as
a comprehensive overview of the subject. We refer readers
to (Taylor 1988; Paul and Kaufman 2014b) for comprehensive
overviews of the different definitions of creativity.
We can describe a person (e.g. artist, poet), a product
(painting, poem), or the mental process as being creative
(Taylor 1988; Paul and Kaufman 2014a). Among the
various definitions of creativity it seems that there is a convergence
to two main conditions for a product to be called
“creative”. That product must be novel, compared to prior
work, and also has to be of value or influential (Paul and
Kaufman 2014a). These criteria resonate with Kant’s definition
of artistic genius, which emphasizes two conditions
“originality” and being “exemplary” 1. Psychologists would
1
Among four criteria for artistic genius suggested by Kant, two
describe the characteristic of a creative product “That genius 1) is
a talent for producing that for which no determinate rule can be
given, not a predisposition of skill for that which can be learned in
accordance with some rule, consequently that originality must be
it’s primary characteristic. 2) that since there can also be original
nonsense, its products must at the same time be models, i.e., exemplary,
hence, while not themselves the result of imitation, they must
not totally agree with this definition since they favor associating
creativity with the mental process that generates the
product (Taylor 1988; Nanay 2014). However associating
creativity with products makes it possible to argue in favor
of “Computational Creativity”, since otherwise, any computer
product would be an output of an algorithmic process
and not a result of a creative process. Hence, in this paper
we stick to quantifying the creativity of products instead of
the mental process that create the product.
Boden suggested a distinction between two notions of
creativity: psychological creativity (P-creativity), which assesses
novelty of ideas with respect to its creator, and historical
creativity (H-creativity), which assesses novelty with
respect to the whole human history (Boden 1990). It follows
that P-creativity is a necessary but not sufficient condition
for H-creativity, while H-creativity implies P-creativity (Boden
1990; Nanay 2014). This distinction is related to the
subjective (related to person) vs. objective creativity (related
to the product) suggested by Jarvie (Jarvie 1986). In this paper
our definition of creativity is aligned with objective/Hcreativity,
since we mainly quantify creativity within a historical
context.
Computational Framework
According to the discussion in the previous section, a creative
product must be original, compared to prior work, and
valuable (influential) moving forward. Let us construct a
network of creative products and use it to assign a creativity
score to each product in the network according to the aforementioned
criteria. In this section, for simplicity and without
loss of generality, we describe the approach based on a
network of paintings, however the framework is applicable
to other art or literature forms.
Constructing a Painting Graph
Let us denote by P = {pi, i = 1 ··· N} a set of paintings.
The goal is to assign a creativity score for each painting, denoted
by C(pi) for painting pi . Every painting comes with
a time label indicating the date it was created, denoted by
t(pi). We create a directed graph where each vertex corresponds
to a painting. A directed edge (arc) connects painting
pi to pj if pi was created before pj . Each directed edge is
assigned a positive weight (we will discuss later where the
weights come from), we denote the weight of edge (pi, pj )
by wij . We denote by Wij the adjacency matrix of the painting
graph, where Wij = wij if there is an edge from pi to
pj and 0 otherwise. Note that according to this definition, a
painting is not connected to itself, i.e., wii = 0, i = 1 ··· N.
By construction, wij > 0 ! wji = 0, i.e., the graph is
anti-symmetric.
To assign the weights we assume that there is a similarity
function that takes two paintings and produces a positive
scalar measure of affinity between them (higher value indicates
higher similarity). We denote such a function by S(·, ·)
yet serve others in that way, i.e., as a standard or rule for judging.”
(Guyer and Wood 2000)-p186
Proceedings of the Sixth International Conference on Computational Creativity June 2015 40
Figure 1: Illustration of the construction of the Creativity Implication Network: blue arrows indicate temporal relation and
orange arrows indicate reverse creativity implication (converse).
and, therefore,
wij =
⇢ S(pi, pj ) if t(pi) < t(pj ).
0 otherwise.
Since there are multiple possible visual aspects that can be
used to measure similarity, we denote such a function by
Sa(·, ·) where the superscript a indicates the visual aspect
that is used to measure the similarity (color, subject matter,
brush stroke, etc.) This implies that we can construct
multiple graphs, one for each similarity function. We denote
the corresponding adjacency matrix by Wa, and the
induced creativity score by Ca, which measure the creativity
along the dimension of visual aspect a. In the rest of this
section, for the sake of simplicity, we will assume one similarity
function and drop the superscript. Details about the
similarity function will be explained in the next section.
Creativity Propagation
Giving the constructed painting graph, how can we propagate
the creativity in such a network? To answer this question
we need to understand the implication of the weight of
the directed edge connecting two nodes on their creativity
scores. Let us assume that initially we assign equal creativity
indices to all nodes. Consider painting pi and consider an
incoming edge from a prior painting pk. A high weight on
that edge (wki) indicates a high similarity between pi and pk,
which indicates that pi is not novel, implying that we should
lower the creativity score of pi (since pi is subsequent to
pk and similar to it) and increase the creativity score of pk.
In contrast, a low weight implies that pi is novel and hence
creative compared to pk, therefore we need to increase the
creativity score of pi and decreases that of pk.
Let us now consider the outgoing edges from pi. According
to our notion of creativity, for pi to be creative it is not
enough to be novel, it has to be influential as well (some
others have to imitate it). This indicates that a high weight,
wij , between pi and a subsequent painting pj implies that
we should increase the creativity score of pi and decrease
that of pj . In contrast, a lower weight implies that pi is not
influential on pj , and hence we should decrease the score for
pi and increase it for pj . These four cases are illustrated in
Figure 1. A careful look reveals that the two cases for the
incoming edges and those for the outgoing edges are in fact
the same. A higher weight implies the prior node is more
influential and the subsequent node is less creative, and a
lower weight implies the prior node is less influential and
the subsequent node is more creative.
Creativity Implication Network
Before converting this intuition to a computational approach,
we need to define what is considered high and low
for weights. We introduce a balancing function on the graph.
Let m(i) denote a balancing value for node i, where for the
edges connected to that node a weight above m(i) is considered
high and below that value is considered low. We define
a balancing function as a linear function on the weights connecting
to each node in the form
Bi(w) = ⇢ w " m(i) if w > 0.
0 otherwise.
We can think of different forms of balancing functions that
can be used. Also there are different ways to set the parameter
m(i) with different implications, which we will discuss
in the next section. This form of balancing function basically
converts weights lower than m(i) to negative values.
The more negative the weight of an edge the more creative
the subsequent node and the less influential the prior node.
The more positive the weight of an edge the less creative the
subsequent node and the more influential the prior node.
The introduction of the negative weights in the graph, despite
providing a solution to represent low weights, is problematic
when propagating the creativity scores. The intuition
is, a negative edge between pi and pj is equivalent to
a positive edge between pj and pi. This directly suggests
that we should reverse all negative edges and negate their
values. Notice that the original graph construction guarantees
that an edge between pi and pj implies no edge between
pj and pi, therefore there is no problem with edge reversal.
This process results in what we call “Creativity Implication
Network”. We denote the weights of that graph by w˜ij and
its adjacency matrix by W
f. This process can be described
Proceedings of the Sixth International Conference on Computational Creativity June 2015 41
mathematically as
B(wij ) > 0 ! w˜ij = B(wij )
B(wij )=0 ! w˜ij = 0
B(wij ) < 0 ! w˜ji = "B(wij )
The Creativity Implication Network has one simple rule that
relates its weights to creativity propagation: the higher the
weight of an edge between two nodes, the less creative the
subsequent node and the more creative the prior node. Note
that the direction of the edges in this graph is no longer related
to the temporal relation between its nodes, instead it
is directly inverse to the way creativity scores should propagate
from one painting to another. Notice that the weights
of this graph are non-negative.
Computing Creativity Scores
Given the construction of the Creativity Implication Network,
we are now ready to define a recursive formula for assigning
creativity scores. We will show that the construction
of the Creativity Implication Network reduces the problem
of computing the creativity scores to a traditional network
centrality problem. The algorithm will maintain creativity
scores that sum up to one, i.e., the creativity scores form
a probability distribution over all the paintings in our set.
Given an initial equal creativity scores, the creativity score
of node pi should be updated as
C(pi) = (1 " ↵)
N + ↵
X
j
w˜ij
C(pj )
N(pj )
, (1)
where 0  ↵  1 and N(pj ) = P
k w˜kj . In this formula,
the creativity of node pi is computed from aggregating a
fraction ↵ of the creativity scores from its outgoing edges
weighted by the adjusted weights w˜ij . The constant term
(1 " ↵)/N reflects the chance that similarity between two
paintings might not necessarily indicate that the subsequent
one is influenced by the prior one. For example, two paintings
might be similar simply because they follow a certain
style or art movement. The factor 1 " ↵ reflects the probability
of this chance. The normalization term N(pj ) for
node j is the sum of its incoming weights, which means that
the contribution of node pj is split among all its incoming
nodes based on the weights, and hence, pi will collect only
a fraction w˜ij/
P
k w˜kj of the creativity score of pj .
The recursive formula in Eq 1 can be written in a matrix
form as
C = (1 " ↵)
N
1 + ↵W C, ff (2)
where W
ff is a column stochastic matrix defined as W
ffij =
w˜ij/
P
k w˜kj , and 1 is a vector of ones of the same size as
C. It is easy to see that since W
ff, C, and 1
N 1 are all column
stochastic, the resulting scores will always sum up to one.
The creativity scores can be obtained by iterating over Eq 2
until conversion. Also a closed-form solution for the case
where ↵ 6= 1 can be obtained as
C⇤ = (1 " ↵)
N (I " ↵W
ff)
"11. (3)
A reader who is familiar with social network analysis literature
might directly see the relation between this formulation
and some traditional network centrality algorithms.
Eq 2 represents a random walk in a Markov chain. Setting
↵ = 1, the formula in Eq 2 becomes a weighted variant
to eigenvector centrality (Borgatti and Everett 2006),
where a solution can be obtained by the right eigenvector
corresponding to the largest eigenvalue of W
ff. The formulation
in Eq 2 is also a weighted variant of Hubbell’s centrality
(Hubbell 1965). Finally the formulation can be seen as an
inverted weighted variant of the Page Rank algorithm (Brin
and Page 1998). Notice that this reduction to traditional network
centrality formulations was only possible because of
the way the Creativity Implication Network was constructed.
Originality vs. Influence
The formulation above sums up the two criteria of creativity,
being original and being influential. We can modify the formulation
to make it possible to give more emphasis to either
of these two aspects when computing the creativity scores.
For example it might be desirable to emphasize novel works
even though they are not influential, or the other way around.
Recall that the direction of the edges in Creativity Implication
Network are no longer related to the temporal relation
between the nodes. We can label (color) the edges in the
network such that each outgoing edge e(pi, pj ) from a given
node pi is either labeled as a subsequent edge or a prior edge
depending on the temporal relation between pi and pj . This
can be achieved by defining two disjoint subsets of the edges
in the networks
Eprior = {e(pi, pj ) : t(pj ) < t(pi)}
Esubseq = {e(pi, pj ) : t(pj ) % t(pi)}
This results in two adjacency matrices, denoted by W
fp and
W
fs such that W
f = W
fp + W
fs, where the superscripts p and s denote the prior and subsequent edges respectively. Now
Eq 1 can be rewritten as
C(pi) = (1 ! ↵)
N + (4)
↵["
X
j
w˜p
ij
C(pj )
Np(pj ) + (1 ! ")
X
j
w˜s
ij
C(pj )
Ns(pj )
],
where Np(pj ) = P
k w˜p
kj and Ns(pj ) = P
k w˜s
kj . The
first summation collects the creativity scores stemming from
prior nodes, i.e., encodes the originality part of the score,
while the second summation collects creativity scores stemming
from subsequent nodes, i.e, encodes influence. We introduced
a parameter 0  "  1 to control the effect of the
two criteria on the result. The modified formulation above
can be written as
C = (1 " ↵)
N
1 + ↵["W
ggpC + (1 " ")W
ggsC], (5)
where W
ggp and W
ggs are the column stochastic adjacency matrices
resulting from normalizing the columns of W
fp and
W
fs respectively. It is obvious that the closed-form solution
Proceedings of the Sixth International Conference on Computational Creativity June 2015 42
in Eq 3 is applicable to this modified formulation where W
ff
is defined as W
ff = "W
ggp + (1 " ")W
ggs.
Creativity Network for Art
In this section we explain how the framework can be realized
for the particular case of visual art.
Visual Likelihood: For each painting we can use computer
vision techniques to obtain different feature representations
for its image, each encoding a specific visual aspect(s) related
to the elements and principles of arts. We denote such
features by fa
i for painting pi, where a denotes the visual
aspect that the feature quantifies. We define the similarity
between painting pi and pj , as the likelihood that painting
pj is coming from a probability model defined by painting
pi. In particular, we assume a Gaussian probability density
model for painting pi, i.e.,
Sa(pj , pi) = P r(pj |pi, a) = N (·; fa
i , #aI).
It is important to limit the connections coming to a given
painting. By construction, any painting will be connected to
all prior paintings in the graph. This makes the graph highly
biased since modern paintings will have extensive incoming
connections and early paintings will have extensive outgoing
connections. Therefore we limit the incoming connections
to any node to at most the top K edges (the K most similar
prior paintings).
Temporal Prior: It might be desirable to add a temporal
prior on the connections. If a painting in the nineteenth century
resembles a painting from the fourteenth century, we
shouldn’t necessarily penalize that as low creativity. This
is because certain styles are always reinventions of older
styles, for example neoclassicism and renaissance. Therefore,
these similarities between styles across distant time periods
should not be considered as low creativity. Therefore,
we can add a temporal prior to the likelihood as
Sa(pj , pi) = P rv(pj |pi, a) · P rt
(pj |pi),
where the second probability is a temporal likelihood (what
is the likelihood that pj is influenced pi given their dates)
and the first is the visual likelihood. There are different ways
to define such a temporal likelihood. The simplest way is
a temporal window function, i.e., P rt
(pj |pi)=1 if pi is
within K temporal neighbors prior to pj and 0 otherwise2.
Balancing Function: There are different choices for the
balancing function B(w), as well as the parameter for that
function. We mainly used a linear function for that purpose.
The parameter m can be set globally over the whole
graph, or locally for each time period. A global m can be
set as the p-percentile of the weights of the graph, which
is p-percentile of all the pairwise likelihoods. This directly
means that p% of the edges of the graph will be reversed
when constructing the Creativity Implication Graph.
2
Alternatively, a Gaussian density can be use, P rt
(pj |pi) =
exp(![t(pi)!t(pj )]2/#2
t ). However, adding such temporal Gaussian
would complicate the algorithm since it will not be easy to
estimate a suitable #t, specially the graph can have non-uniform
density over the time line.
One disadvantage of a global balancing function is that different
time periods have different distributions of weights.
This suggests using a local-in-time balancing function. To
achieve that we compute mi for each node as p% of the
weight distribution based on its temporal neighborhood.
Experiments and Results
Datasets and Visual Features
Artchive: This dataset was previously used for style classification
and influence discovery (Saleh et al. 2014). It
contains a total of 1710 images of art works (paintings and
sculptures) by 66 artists, from 13 different styles from 1412-
1996, chosen from Mark Harden’s Artchive database of fineart
(Harden ). The majority of the images are of the full
work, while a few are details of the work.
Wikiart.org: We used the publicly available dataset of
“Wikiart paintings”3; which, to the best of our knowledge,
is the largest online public collection of artworks. This collection
has images of 81,449 fine-art paintings and sculptures
from 1,119 artist spanning from 1400-2000+. These
paintings are from 27 different styles (Abstract, Byzantine,
Baroque, etc.) and 45 different genres (Interior, Landscape,
Portrait, etc.). We pruned the dataset to 62,254 western
paintings by removing genres and mediums that are not suitable
for the analysis such as sculpture, graffiti, mosaic, installation,
performance, photos, etc.
For both datasets the time annotation is mainly the year.
Therefore, it is not possible to tell which is prior between
any pair of paintings with the same year of creation. Therefore
no edge is added between their corresponding nodes.
We experimented with different state-of-the-art feature
representations. In particular, the results shown here are using
Classeme features (Torresani, Szummer, and Fitzgibbon
2010). These features were shown to outperform other stateof-the-art
features for the task of style classification (Saleh
et al. 2014). These features (2659 dimensions) provide
semantic-level representation of images, by encoding the
presence of a set of basic-level object categories (e.g. horse,
cross, etc.), which captures the subject matter of the painting.
Some of the low-level features used to learn the
Classeme features also capture the composition of the scene.
Example Results
We show qualitative and quantitative experimental results of
the framework applied to the aforementioned datasets. As
mentioned in the introduction, any result has to be evaluated
given the set of paintings available to the algorithm and the
capabilities of the visual features used. Given that the visual
features used are mainly capturing subject matter and
composition, sensible creativity scores are expected to re-
flect these concept. A low creativity score does not mean
that the work is not creative in general, it just means that the
algorithm does not see it creative with respect to its encoding
of subject matter and composition.
Figures 2-top and 3 show the creativity scores obtained
on the Artchive and Wikiart datasets respectively. Figure 2-
bottom shows a zoom in to the period between 1850-1950 in
3
http://www.wikiart.org/
Proceedings of the Sixth International Conference on Computational Creativity June 2015 43
the Artchive dataset, which is very dense in the graph4. In all
figures we plot the scores vs. the year of the painting. The
figures visualize some of the paintings that obtained high
scores, as well as some with low scores (the scores in the
plots are scaled). We randomly sampled points with low
scores for visualization. A close look at the paintings that
scored low (bottom) reveals the presence of typical subject
matter that is common in the dataset, or in some cases the
image presents an unclear view of a sculpture (e.g. Rodin
1889 sculpture in the bottom right). The general trend shows
peaks in creativity around the time of High Renaissance (late
15th , early 16th century) and the late 19th and early 20th
centuries, and a significant increase in the second half of the
20th century.
One of the interesting findings is the algorithm’s ability to
point out wrong annotations in the dataset. For example, one
of the highest scoring paintings around 1910 was a painting
by Piet Mondrain called “ Composition en blanc, rouge et
jaune,” (see Figure 2). By examining this painting, we found
that the correct date for it is around 1936 and it was mistakenly
annotated in the Artchive dataset as 1910 5. Modrain
did not start to paint in this grid-based (Tableau) style untill
around 1920. So it is no surprise that wrongly dating one
of Mondrain’s tableau paintings to 1910 caused it to obtain
high creativity score, even above the cubism paintings from
that time. On the Wikiart dataset, one of the highest-scored
painting was “tornado” by contemporary artist Joe Goode,
which was found to be mistakenly dated 1911 in Wikiart 6.
A closer look at the artist biography revealed that he was
born in 1937 and this painting was created in 19917. It
is not surprising for a painting that was created in 1991 to
score very high in creativity if it was wrongly dated to 1911.
These two examples, besides indicating that the algorithm
works, show the potential of proposed algorithm in in spotting
wrong annotations in large datasets, which otherwise
would require tremendous human effort.
Time Machine Experiment
Given the absence of ground truth for creativity, the aforementioned
wrong annotations inspired us with a methodology
to quantitatively evaluate the framework. We designed
what we call “time machine” experiment, where we change
the date of an artwork to some point in the past or some point
in the future, relative to its correct time of creation. Then we
compute the creativity scores using the wrong date, by running
the algorithm on the whole data. We then compute the
gain (or loss) in the creativity score of that artwork compared
to its score using correct dating. What should we expect
4
For Figure 2 a temporal window historical prior is uses.
For Figure 3 no historical prior was used. For both, we set
K=500,↵=0.15 5
The wrong annotation is in the Artchive CD obtained in 2010.
The current online version of Artchive has corrected annotation for
this painting 6
http://www.wikiart.org/en/search/tornado/
1#supersized-search-318512 - accessed on Feb 28th,
20157
http://www.artnet.com/artists/joe-goode/
tornado-9-2Y7erPME95YlkhFp7DRWlA2
Table 1: Time Machine Experiment
Art movement avg % gain/loss % increase
Moving backward to AD 1600
Neoclassicism 5.78%±1.28 97%±4.8
Romanticism 7.52%± 2.04 98%± 4.2
Impressionism 14.66%± 2.78 99%±3.2
Post-Impressionism 16.82%±2.22 99%±3.1
Symbolism 15.2%±2.94 97%±4.8
Expressionism 16.83%±2.43 98%±4.2
Cubism 13.36%±2.43 89%±9.9
Surrealism 12.66%±1.82 95%±7.1
American Modernism 11.75%±2.99 84%±8.4
Wandering around to AD 1600
Renaissance 0.68 %± 2.05 39%±5.7
Baroque 2.85%± 1.09 71%±19.7
Moving forward to AD 1900
Renaissance -8.13%± 2.02 20%±10.5
Baroque -10.2%±2.03 0%±0
from an algorithm that assigns creativity in a sensible way?
Moving a creative painting back in history would increase
its creativity score, while moving a painting forward would
decrease its creativity. Therefore, we tested three settings:
I) Moving back to AD 1600: For styles that date after 1750,
we set the test paintings back to a random date around 1600
using Normal distribution with mean 1600 and std 50 years,
i.e. N(1600, 502) . II) Moving forward to AD 1900: For the
Renaissance and Baroque styles, we set the test paintings
to random dates around 1900 sampled from N(1900, 502).
III) Wandering about AD 1600 (baseline): In this experiment,
for the Renaissance and Baroque styles, we set the
test paintings to random dates around 1600 sampled from
N(1600, 502).
Table 1 shows the results of these experiments. We ran
this experiment on the Artchive dataset with no temporal
prior. In each run we randomly selected 10 test paintings
of a given style and applied the corresponding move. We
used 10 as a small percentage of the data set (less than 1%),
not to disturb the global distribution of creativity. We repeated
each experiment 10 times and reported the mean and
standard deviations of the runs. For each style we computed
the average gain/loss of creativity scores by the time
move. We also computed the percentage of the test paintings
whose scores have increased. From the table we clearly see
that paintings from Impressionist, Post-Impressionist, Expressionist,
and Cubism movements have significant gain in
their creativity scores when moved back to 1600. In contrast,
Neoclassicism paintings have the least gain, which makes
sense, because Neoclassicism can be considered as revival
to Renaissance. Romanticism paintings also have a low gain
when moved back to 1600, which is justified because of the
connection between Romanticism and Gothicism and Medievalism.
On the other hand, paintings from Renaissance
and Baroque styles have loss in their scores when moved
forward to 1900, while they did not change much in the
wandering-around-1600 setting.
Proceedings of the Sixth International Conference on Computational Creativity June 2015 44
1400 1500 1600 1700 1800 1900 2000
1
1.05
1.1
1.15
1.2
1.25
1.3
1.35
1.4
1.45
1.5
Mantegna; 1454
Leonardo; 1469
Mantegna
1474
Velazquez
1632
Vermeer
 1661
Michelangelo
1564
Durer; 1503 Durer; 1514
Goya
1780 Vermeer
1664
Ingres; 1806 Ingres; 1851 Ingres; 1852
Monet
 1865
Rodin; 1889
Munch
 1893
Rousseau; 1898
Mondrian
 1910
Lichtenstein; 1972
1850 1860 1870 1880 1890 1900 1910 1920 1930 1940 1950
1
1.05
1.1
1.15
1.2
1.25
1.3
1.35
Cezanne; 1887
Van Gogh
1889
Rousseau
1897
Klimt
1898
Picasso
1903
Picasso; 1907
Mondrian
1910
Picasso; 1912
Picasso; 1912
Braque
1913
Malevich
1916
Malevich; 1915
Okeeffe
1919
Mondrian
1918 Malevich; 1929
Degas; 1870
Pissarro; 1885 Rodin; 1889 Cezanne; 1906
Figure 2: Top: Creativity scores for 1710 paintings from Artchive dataset. Bottom: zoom in to the period 1850-1950. Each
point represents a painting. The thumbnails illustrate some of the paintings that scored relatively high or low compared to their
neighbors. Only artist names and dates of the paintings are shown on the graph because of limited space. The red-dotted-framed
painting by Piet Mondrain scored very high because it was wrongly dated to 1910 instead of 1936 in the dataset.
Proceedings of the Sixth International Conference on Computational Creativity June 2015 45
1400 1500 1600 1700 1800 1900 2000 2100
1
2
3
4
5
6
7
8
9
10
Rubens
1635
Gozzoli
1461
Da Vinci; 1480
Signorelli
 1482
Carracci
1605
Dorazio
 1998
Calhau
 1998
Dorazio; 1990
Lefranc
 1929
Gilliam; 1965
Mantegna;1502 Veronese;1558 Winterhalter Kittelsen; 1911 Reni, 1625 Canaletto; 1725 Fragonard; 1778 ; 1860
Berkowitz
 1978
Landfield
 1968
Pissarro; 1897
Spilliaert
1908
Mattis; 1916
Bruegel
1568
Durer
1497
Cranach; 1536
Figure 3: Creativity scores for 62K paintings from the wikiart.org dataset
Conclusion and Discusion
The paper presented a computational framework to assess
creativity among a set of products. We showed that, by constructing
a creativity implication network, the problem reduces
to a traditional network centrality problem. We realized
the framework for the domain of visual art, where
we used computer vision to quantify similarity between artworks.
We validated the approach qualitatively and quantitively
on two large datasets.
In this paper we focused on “creative” as an attribute
of a product, in particular artistic products such as painting,
where creativity of a painting is defined as the level of
its originality and influence. However, the computational
framework can be applied to other forms such as sculpture,
literature, science etc. Quantifying creativity as an attribute
of a product facilitates quantifying the creativity of the person
who made that product, as a function over the creator’s
set of products. Hence, our proposed framework also serves
as a way to quantify creativity as an attribute for people.
<references_biblio/>
References
Boden, M. A. 1990. The creative mind: Myths and mechanisms.
Basic Books.
Borgatti, S. P., and Everett, M. G. 2006. A graph-theoretic
perspective on centrality. Social networks 28(4):466–484.
Brin, S., and Page, L. 1998. The anatomy of a large-scale
hypertextual web search engine. Computer networks and
ISDN systems 30(1):107–117.
Fichner-Rathus, L. 2008. Foundations of Art and Design.
Clark Baxter.
Guyer, P., and Wood, A. W. 2000. Critique of the Power
of Judgement. The Cambridge Edition of the Works of Immanuel
Kant. Cambridge University Press.
Harden, M. The artchive@http://artchive.com/cdrom.htm.
Hubbell, C. H. 1965. An input-output approach to clique
identification. Sociometry 377–399.
Jarvie, I. 1986. The rationality of creativity. In Thinking
about Society: Theory and Practice. Springer. 282–301.
Nanay, B. 2014. An experiential account of creativity. In
Paul, E. S., and Kaufman, S. B., eds., The Philosophy of
Creativity: New Essays. Oxford University Press.
Paul, E. S., and Kaufman, S. B. 2014a. Introducing the
philosophy of creativity. In The Philosophy of Creativity:
New Essays. Oxford University Press.
Paul, E. S., and Kaufman, S. B. 2014b. The Philosophy of
Creativity: New Essays. Oxford University Press.
Saleh, B.; Abe, K.; Arora, R. S.; and Elgammal, A. 2014.
Toward automated discovery of artistic influence. Multimedia
Tools and Applications - Special Issue on Multimedia for
Cultural Heritage.
Spratt, E. L., and Elgammal, A. 2014. Computational
beauty: Aesthetic judgment at the intersection of art and
science. In ECCV 2014 Workshops, Part I, Proceedings of
when vision meets art (VisArt) workshop, Lecture Notes on
Computer Science number 8925. Springer.
Taylor, C. W. 1988. Various approaches to and definitions
of creativity. The nature of creativity 99–121.
Torresani, L.; Szummer, M.; and Fitzgibbon, A. 2010. Effi-
cient object category recognition using classemes. In ECCV.
Proceedings of the Sixth International Conference on Computational Creativity June 2015 46