Representing Beauty: Towards a Participatory but Objective Latent Aesthetics

1. Problem Statement

The paper addresses the philosophical and technical challenge of defining "beauty" within machine learning. While beauty is often viewed as culturally constructed, subjective, or elusive, deep learning systems have demonstrated an increasing ability to predict aesthetic judgments. The core problem is determining whether this capability implies:

• Constructivism: Machines are merely mimicking socially constructed (e.g., Western-centric) biases found in training data.
• Realism: There is an objective, physical basis for aesthetic structure that machines can discover independently of human annotation.

The authors aim to resolve the tension between the formal diversity of beautiful objects (mountains, acts of kindness, paintings) and the perceptual centrality of beauty, investigating whether neural networks converge on a "universal representation" of beauty that transcends specific datasets or modalities.

2. Methodology

The authors employ a multi-faceted approach combining representational geometry analysis, statistical testing, and philosophical framing:

• Cross-Model Representational Alignment:

  • The study utilizes the Mutual Nearest Neighbors (MNN) metric to measure representational alignment between a reference model (DINOv2-Large) and various other models with distinct architectures, training objectives, and data modalities (e.g., CLIP, MobileCLIP, ChineseCLIP, VICReg).
  • Dataset: Images are sourced from the Aesthetic Visual Analysis (AVA) dataset, categorized as "aesthetic" (beautiful) or "unaesthetic."
  • Hypothesis: If beauty has a realist basis, models trained on different data should produce highly aligned internal representations for beautiful images, but not for unaesthetic ones.

• Control for Data Congruence (Domain Bias):

  • To rule out the critique that alignment is caused by "in-domain" familiarity (e.g., Western bias in training data), the authors use Maximum Mean Discrepancy (MMD).
  • They compare the distribution of aesthetic vs. unaesthetic AVA images against a subset of ImageNet.
  • Semantic Disentanglement: To control for semantic clusters (e.g., if all beautiful images happen to be of cats), the authors use a ResNet classifier to ensure that aesthetic and unaesthetic subsets contain an equal number of unique semantic classes (one image per class), effectively isolating the "aesthetic" variable from "subject matter."

• Cross-Cultural Validation:

  • The study includes ChineseCLIP, a model trained almost entirely on Chinese data, to test if the convergence of representations holds across non-Western cultural contexts.

• Layer-Wise Analysis:

  • The authors analyze alignment across different network layers to observe the hierarchy of abstraction, testing whether middle layers (where abstract concepts form) show higher alignment for aesthetic content.

3. Key Contributions

• Empirical Evidence for Realist Aesthetics: The paper provides data suggesting that aesthetic content produces significantly higher representational alignment across diverse models than unaesthetic content. This implies a "realist basis" for beauty, challenging the notion that aesthetics are purely socially constructed or arbitrary.
• Refinement of the "Platonic Representation Hypothesis" (PRH): The authors argue that the convergence of representations is better described by Aristotelian Hylomorphism (the unity of matter and form) rather than Platonism. They posit that "ideal" forms are not detached metaphysical entities but are implicit in the physical world and learned through the constraints of natural environments and human perception.
• Identification of Beauty as a "Teleological Attractor": The paper proposes that beauty acts as a unifying principle (a "binder") in latent spaces. Because human culture implicitly curates content toward beauty (a telos), deep learning systems trained on human-generated data naturally converge on this structure.
• Methodological Controls: The study rigorously controls for semantic bias and cultural origin (using ChineseCLIP and semantic disentanglement), strengthening the claim that the observed convergence is driven by the intrinsic structure of beauty rather than dataset artifacts.

4. Results

• Higher Alignment for Aesthetic Images: Figure 1 and Figure 2 demonstrate that models trained on distinct data (including self-supervised unimodal models) show significantly higher representational alignment for aesthetic images compared to unaesthetic or ambiguous images.
• Cross-Cultural Convergence: The inclusion of ChineseCLIP confirms that this alignment is not exclusive to Western-trained models, suggesting the phenomenon is robust across different cultural training corpora.
• Rejection of "In-Domain" Bias: The MMD analysis (Figure 3) reveals that aesthetic images are not more "in-domain" (closer to ImageNet distributions) than unaesthetic images; in fact, unaesthetic images were slightly more in-domain. This refutes the argument that alignment is merely a result of models recognizing familiar Western-style images.
• Semantic Disentanglement: Even when controlling for semantic clusters (ensuring equal class diversity), the aesthetic stratification in cross-model convergence persisted (Figure 5). This confirms that the alignment is driven by aesthetic qualities, not just subject matter.
• Layer-Wise Abstraction: The analysis shows that middle layers of the networks exhibit the highest alignment for aesthetic content, supporting the hypothesis that beauty is an abstract, higher-order concept emerging from the processing of particulars.

5. Significance and Implications

• Philosophical Shift: The paper challenges post-structuralist and constructivist views in digital humanities, arguing for an objective, human-mediated structure to beauty. It suggests that beauty is a manifestation of order and harmony inherent in reality, which machines can detect.
• Human-Machine Co-Creation: The findings reframe AI not as a "creative parrot" mimicking human taste, but as a tool capable of uncovering latent order in cultural and physical reality. Because machines can process data at a scale humans cannot, they can offer novel insights into the structure of beauty.
• Future of Art and AI: The authors propose a symbiotic framework where human creativity (the projection of form) and machine perception (the apprehension of universal geometry) work together. This moves the discourse from "AI replacing artists" to "human-machine co-cultivation of the beautiful."
• Theoretical Framework: By adopting Hylomorphism, the paper offers a robust philosophical grounding for explainable AI (XAI) in aesthetics, suggesting that the "form" learned by AI is inextricably linked to the "matter" of the physical world and human biological/cultural constraints.

In conclusion, the paper argues that beauty is not merely a subjective cultural construct but a teleological attractor in latent space. Deep learning models, trained on human-curated data, converge on this structure, revealing a shared, objective geometry of beauty that bridges human perception and machine intelligence.