Artificial intelligence models can spontaneously develop an ability to recognize human facial beauty without any prior training or exposure to human preferences. A recent study published in Psychology of Aesthetics, Creativity, and the Arts suggests that this aesthetic sense emerges naturally from the way computer networks process basic visual information. The research hints that human aesthetic preferences might also have an innate origin rooted in our basic sensory wiring.
For centuries, philosophers and scientists have debated whether humans are born with an innate sense of beauty or if aesthetic preferences are learned through cultural exposure. In human infants and nonhuman primates, researchers have observed a natural preference for faces that adults deem attractive. It remains difficult to determine whether this preference is hardwired into the brain before birth or acquired rapidly through early visual experiences.
Deep neural networks, a type of artificial intelligence designed to mimic the human brain’s visual processing pathways, offer a way to test this question. In an untrained state, these networks contain randomized connections and have no prior visual experience or knowledge of human preferences. If an untrained network can detect beauty, it suggests the ability is a byproduct of basic structural encoding rather than a learned behavior.
Researchers Tianxin Shu and Delong Zhang from South China Normal University, along with their colleagues, designed an experiment to explore this phenomenon. They aimed to see if artificial neural units would naturally respond to varying degrees of facial beauty. The research team focused on understanding how simple sensory properties might inherently shape aesthetic perception.
The researchers first generated a dataset of human faces using a specialized artificial intelligence program capable of producing highly realistic images. They manipulated the underlying code of these artificial faces to create different levels of beauty. A small group of human volunteers assessed the images to ensure they still looked like natural human faces.
This process yielded more than 10,000 facial images with distinctly graded aesthetic values. The researchers fed these images into an untrained version of a deep neural network called AlexNet. This network consists of multiple layers of artificial neurons that process visual information, starting with simple edges and moving to whole objects.
Because the network was untrained, its internal weights, which are the mathematical rules determining how neurons interact, were completely random. The researchers analyzed the responses of artificial neurons in the network’s final visual processing layer. They discovered that specific neural units spontaneously began responding selectively to different levels of facial beauty.
These “beauty-selective units” appeared consistently, regardless of the face’s orientation or the specific random initialization of the network. To understand how the network processed this information, the researchers altered the facial images. They applied random scrambling, which disrupted the overall shape of the face but kept local details intact.
They also utilized a different type of scrambling that preserved the general spatial arrangement but obscured fine details. Additionally, they tested images reduced entirely to their simple outlines. The team found that both the specific facial features and the spatial arrangement of those features were necessary for the network to evaluate beauty.
In psychological research, this is known as dual-code theory. This theory suggests that humans rely on both the detailed traits of a face and its broader geometric layout to recognize it. The neural network processed this visual information hierarchically, much like the human visual system.
The network’s earlier layers handled simple geometric features and local details. The later layers then integrated this fragmented data into a complete representation of facial beauty. The research team then tested whether these beauty-selective units could actively compare faces.
They trained a simple sorting algorithm to use the responses from these specific units to judge which of two faces was more attractive. The beauty-selective units successfully differentiated between varying levels of beauty. They performed better than units that simply processed raw image pixels.
The accuracy of these units increased as the difference in beauty between the two faces grew larger. Next, the scientists investigated how the final layer of the network arrived at its beauty preferences. They tracked the activity of units in the earlier layers of the network.
They identified specific neurons that responded to beauty on a sliding scale. The activity of these units steadily increased or steadily decreased as the faces became more attractive. The researchers mapped how these earlier units connected to the final beauty-selective units.
They found that the network constructed its sense of beauty by combining the inputs from these earlier, simpler neurons. Units that preferred highly attractive faces received stronger signals from neurons that increased their activity in response to beauty. Finally, to ensure their results were not unique to one specific network or dataset, the researchers repeated their experiments.
They used a different neural network architecture, known as VGG-16, and again found that beauty-selective units emerged spontaneously. They also tested the network using a database of 200 real human faces spanning different racial backgrounds. The beauty-selective units successfully evaluated the attractiveness of these natural faces.
This outcome proved their ability generalized beyond artificial images. While these findings offer an intriguing look at artificial cognition, a network’s sense of beauty is not identical to human aesthetic appreciation. The study focused on a mathematical abstraction of facial structure rather than the subjective emotional experience humans associate with beauty.
A generalized visual preference for symmetrical or average features does not equate to the rich psychological experience of finding someone beautiful. The artificial faces used in the primary experiments were heavily standardized, minimizing variations in lighting, background, and hairstyle. Real-world aesthetic judgments involve highly variable environments and subjective personal preferences that these models cannot capture.
The current algorithm also primarily evaluated a single baseline standard of beauty. Future research could explore how untrained networks respond to other aesthetic categories, such as natural landscapes, architecture, or fine art. Expanding the criteria for beauty could help scientists determine if this inherent perceptual bias applies to all visual structures or just human faces.
The study, “The Spontaneous Emergence of “A Sense of Beauty ” in Untrained Deep Neural Networks,” was authored by Tianxin Shu, Huawei Xu, Xingxing Chen, Yuxuan Cai, Ming Liu, and Delong Zhang.