Can Artificial Intelligence (AI) tell us when “a rose by any other word smells as sweet”?

Published in

ITNEXT

9 min readJan 11, 2018

Click here to share this article on LinkedIn »

“Can “a rose by any other word smell as sweet”?” Can a thing of beauty, by any other quality be as beautiful? That’s a question we may be asking ourselves more often as Artificial Intelligence (AI) applications, like those offered by Paralleldots Inc., begin to tell us what is and is not beautiful.

Shashank Gupta and his colleagues at Paralleldots Inc. recently announced they’ve developed a Deep Convolutional Neural Network (CNN) which can be “trained to recognize an image’s “aesthetic quality”. They’ve provided a demo of how their CNN for visual analysis can be applied at this website: ( https://www.paralleldots.com/visual-analytics )

Convolutional Neural Network architecture with dense blocks and feature accumulation from different levels to model aesthetics

The Paralleldots team trains a model from scratch to get accuracy of 78.7% on AVA2 Dataset close to the best models available (85.6%). They further show accuracy increases to 81.48% by increasing the training set to 10 percentile of entire AVA dataset, to demonstrate their algorithm gets better with more data.

According to the Paralleldots team their Virality Detection Application Programming Interface (API) gives a score to images based on their potential to go viral on the Internet. The Virality Detection API predicts the popularity of photos based on similar, trending photos on social media. While this is not the same as the beauty attributed to an image because of any objective characteristics, it is quite similar to Immanuel Kant’s condition of “universality” which he used as one measure of beauty.

According to the Paralleldots researchers, “Visual aesthetic analysis is the task of classifying images into being perceived as attractive or unattractive by humans.” While previously handcrafted image features were a common way to solve aesthetic analysis, CNNs have recently been used to solve the problem as state of the art approaches. They have multiple advantages such as having lesser inference time and the ability to be deployed in mobile devices. A variety of standard CNN architectures pretrained on the ImageNet dataset are readily available as open source for use. ImageNet is an image database organized according to the WordNet hierarchy, in which each node of the hierarchy is depicted by hundreds and thousands of images.

Following is a sample of the Google AVA dataset. Researchers at Paralleldots, Inc. used these objects along with hundreds of others in the Google AVA dataset to enable their CNN to perform “visual aesthetic analysis”.

The Virality Detection API was built by training a super deep neural network on a huge corpus of images and their scores crawled from the open web. The score that you get as output is the virality score of the input photo out of maximum score of 100. The Paralleldots team says their in-house experiments suggest that accuracy of their algorithm to predict image virality is as high as 85%. Following are virality scores for the 10 most famous Paintings In the world where the maximum score is 100%.

Painting & Artist — Virality Detection Score

The Scream — Edvard Munch — 95%

2. Guernica — Pablo Picasso — 77%

3. The Last Supper — Leonardo da Vinci — 72%

4. The Creation Of Adam — Michelangelo — 71%

5. The Night Watch — Rembrandt van Rijn — 69%

6. Girl With A Pearl Earring — Johannes Vermeer — 68%

7. The Persistence Of Memory — Salvador Dali — 49%

8. Starry Night — Vincent van Gogh — 43%

9. Self-Portrait Without Beard — Vincent van Gogh — 22%

10. Mona Lisa — Leonardo da Vinci. — 9%

Almost three hundred years before the Paralleldots team developed their CNN to find beauty, the great German philosopher, Immanuel Kant, found that artwork is beautiful insofar as it instigates an intellectual activity termed reflective judgment. According to Kant, “…Beauty is never experienced as a determinate thing. We do not experience beauty directly, although it is always implicated in our experiences of the world. Beauty is a feeling induced by our sense of an ordering, a valuing, at work in the world that lies beyond any explicit demonstration.”

Other conditions may also contribute to what it is to be a judgment of taste, but they are consequential on, or predicated on, two fundamental conditions. In this respect Kant followed the lead of David Hume and other writers in the British sentimentalist tradition (Hume 1757). Kant isolated two fundamental necessary conditions for a judgment to be a judgment of taste (Kant 1790):

subjectivity and
universality (virality)

The first necessary condition of a judgment of taste is that it is essentially subjective. What this means is that the judgment of taste is based on a feeling of pleasure or displeasure. It is this that distinguishes a judgment of taste from an empirical judgment. Central examples of judgments of taste are judgments of beauty and ugliness. They can be about art or nature.

For Kant, the judgment of taste also claims “universal validity”, which he described as demanding or requiring agreement from others in a way we do not in our judgments about the niceness of something like wine, which is just a question of individual preference. In matters of taste and beauty, we think that others ought to share our view. It is because the judgment of taste has such an aspiration to universal validity that it seems “as if [beauty] were a property of things.”

In many ways the Paralleldots team adopted Kant’s condition of universality in their CNN. They simply substituted the modern-day concept of “virality” for what Kant called “universality”. According to the Paralleldots team their Virality Detection Application Programming Interface (API) gives a score to images based on their potential to go viral on the Internet. The Virality Detection API predicts the popularity of photos compared to the “properties” of, trending photos on social media. While this is not the same as the beauty attributed to an image because of objective characteristics, it is quite similar to Kant’s condition of universality.

Eleven hundred years before Kant and Hume the Greeks taught that “Beauty has rules” and rules have to be learned. The most famous rule related to beauty was the Golden Number ( 1.61803399 ) or Golden Ratio which can be stated, using the following graphic to the below, ( a+b is to a as a is to b ). The golden ratio was discovered by the Greek philosopher and mathematician, Pythagoras and is found in many works of art as well as throughout nature.

In the golden ratio, a+b is to a as a is to b.

The golden number is usually represented by the Greek letter ‘Phi’, or Φ, after Phidias, (480–430 bc) the Greek sculptor who used it in the proportions of many of his famously beautiful sculptures. The Parthenon in Athens, built by the ancient Greeks from 447 to 438 BC, is regarded by many to illustrate the application of the Golden Ratio in design. The golden number, defines a harmony of proportions that, according to the Greeks, like Phidias, inspires reverence and admiration of what has been created and a desire to create more beauty on the part of the observer.

Symmetry (from Greek for “agreement in dimensions, due proportion, arrangement”) in everyday language refers to a sense of harmonious and beautiful proportion and balance. In mathematics, “symmetry” has a more precise definition, that an object is invariant to any of various transformations; including reflection, rotation or scaling. Although these two meanings of “symmetry” can be told apart, they are very much related in the concepts of harmony and balance. The Mona Lisa, the Parthenon and a nautilus shell are all applications of the Golden Ratio, a mathematical expression that has long been associated with beauty.

“The history of art shows that in the long search for an elusive cannon of perfect proportion, one that would somehow automatically confer aesthetically pleasing qualities on all works of art, the golden Ratio, has proven to be the most enduring”. ( Mario Livio, The Golden Ratio )

In a ‘beautiful’ face, for example, the distance from the middle of the nose is 1.618’th of the width of the eye. Another example is the distance from the chin to the pupil is 1,618’th of the distance from the pupil to the hairline. In the body, the distance from the sole of the foot to the waist is 1,618’th of the distance from the waist to the crown. This pattern repeats itself over and over in what are regarded as “beautiful” people.

The closer a given subject is to the ratio, the more beautiful it is generally considered to be. Leonardo da Vinci’s Mona Lisa had a number of golden rectangles throughout. By drawing a rectangle around her face, we can see she is indeed golden. If we divide that rectangle with a line drawn across her eyes, we get another golden rectangle, meaning that the proportion of her head length to her eyes is golden. There are other golden rectangles that can be drawn on the rest of her body like from the her neck to the top of her hands.

In a very recent study Nathan Kondamuri, of the Art Institute of MIT, took measurements using two sculptures found in the Hayden Memorial Library at MIT. He also conducted a survey asking subjects to compare the two sculptures and rate them on a scale of 1–10 on their beauty. Upon comparing the measurements and results from this survey, Kondamuri concluded “there is a direct correlation between the golden ratio and aesthetic beauty”.

So while the Mona Lisa may have been seen as beautiful, using an objective standard like the golden ratio, shockingly, she is considered less beautiful using 21st century Artificial Intelligence software and a standard of “universality”.

We may need to question Kant’s idea of requiring agreement from others as well as the Paralleldots Virality Detection score as mechanisms to judge beauty, at least as it relates to generational differences.

During his lifetime, (1912–1956) Jackson Pollock enjoyed considerable fame and notoriety as a major artist of his generation. Following is his famous “Convergence” painting which scored higher using the Virality Detection Application than seven of the ten most famous paintings in the world.

Also following is Henri Matisse’s “Woman with a Hat” which scored an amazing 96% using the Viral Detection Software, again indicating that more contemporary paintings appear more beautiful than paintings from an earlier period. If that’s the case then perhaps, “beauty” as determined by AI software, like Paralleldot’s, is more dependant on the vintage of the dataset used to identify the characteristics of beauty than any objective standard like the Golden Ratio.