“The Spontaneous Emergence of ‘A Sense of Beauty’ in Untrained Deep Neural Networks”, 2024 (; similar):
[see also et al 2021] The sense of facial beauty has long been observed in both infants and nonhuman primates, yet the neural mechanisms of this phenomenon are still not fully understood.
The current study employed generative neural models [StyleGAN 2] to produce facial images of varying degrees of beauty and systematically investigated the neural response of untrained deep neural networks (DNNs) to these faces.
Representational neural units for different levels of facial beauty are observed to spontaneously emerge even in the absence of training. Furthermore, these neural units can effectively distinguish between varying degrees of beauty. Additionally, the perception of facial beauty by DNNs relies on both configuration and feature information of faces. The processing of facial beauty by neural networks follows a progression from low-level features to integration. The tuning response of the final convolutional layer to facial beauty is constructed by the weighted sum of the monotonic responses in the early layers.
These findings offer new insights into the neural origin of the sense of beauty, arising the innate computational abilities of DNNs.
[Keywords: esthetic neurocomputation, generative neural models, untrained deep neural networks, facial beauty, linear weighted summation]
…In conclusion, our research explored the neural origin of beauty sense from the innate computational abilities of DNNs. We found that there are units selectively responsive to facial beauty in the completely randomly initialized DNNs, and the responses of these units are linearly distributed. Additionally, representations of beauty have emerged in the initial layers of DNNs. Untrained DNNs perceive facial beauty in a hierarchical manner, where both configuration information and feature information of faces contribute to the completion of esthetic processing. The selective responses in the final layer of the DNN are constructed through linear weighting of monotonic responses in the early layers.