Louise Bourgeois’s Spiders

AI-Trained Virtual Exhibition Exploring Artist Style and Machine Aesthetics

Role

AI Trainer & Virtual Exhibition Designer

Tool

Unity, Visual Studio, AI Training, Machine Learning

Team

Becky Wang

Project Overview

Louise Bourgeois’s Spiders is a virtual exhibition that explores how AI can learn and recreate an artist’s visual language. By training a model on Bourgeois’s work, the project reimagines her iconic spider forms in an interactive digital space—blending curation, machine aesthetics, and creative coding into one immersive experience.

Harvard SCI 6487 Machine Aesthetics: The Surrogate of Taste

#AI_generated_art #virtual_exhibition #digital_curation #artist_style_replication #machine_aesthetics #interactive_media #computational_creativity

A Machine Aesthetic Inquiry

"Can AI Simulate Artistic Taste?"

This project explores how computational processes might approximate an artist’s aesthetic decisions.

Modeling Aesthetic Judgment

01. The Premise

AI-generated design today often relies on large, centralized models where design intent is reduced to a prompt. But language alone is insufficient for expressing the nuance of aesthetic judgment, especially in visual arts and architecture.

This project questions whether a machine can do more than imitate. Can it learn to recognize what an artist might find meaningful? Can it simulate taste, not as a generic metric, but as a deeply subjective and emotional pattern?

By using Louise Bourgeois’s works as both dataset and conceptual anchor, we trained a model to act as a surrogate of taste, encoding the visual, emotional, and symbolic DNA of her world into a computational form.

Personal Trauma as Aesthetic Logic

02. About Louise Bourgeois

Louise Bourgeois (1911–2010) was a French-American artist whose sculptures, installations, and drawings constructed a deeply personal visual world. Her work draws from childhood memory and domestic life, centering on themes of abandonment, betrayal, family, femininity, and repair.

At the core of her practice is a confrontation with emotional pain—often stemming from her early experiences: her father's long-term affair with the family's live-in English tutor, her mother's silent endurance, and her death when Bourgeois was 22. These private traumas, layered with the historical backdrop of World War I, became the psychological raw material of her life’s work.

“The subject of pain is the business I am in – to give meaning and shape to frustration and suffering.”

Aesthetic Optimization through Surrogate Models

03. From Cell to Spider

This project reframes machine learning as a process of aesthetic optimization, where artistic taste is not inferred through prompts, but embedded into the model itself.

We began with a curated dataset representing Louise Bourgeois’s sculptural preferences:

DATA_like included her documented works: sculptures characterized by raw materiality, psychological tension, and formal restraint.
‍DATA_dislike included visual opposites: soft, decorative, or commercial representations of spiders that contradict her aesthetic values.

Using Python and PyTorch, we trained the classifier on a contrastive image dataset.
Once trained, we used a visualization script to generate a PDF summary showing how the model responds to new images, ranking them based on how strongly they align with the learned aesthetic. This allowed us to see, in visual terms, what the model considered a “like” or “dislike,” according to Bourgeois’s inferred taste.

The image shows how our trained classifier responds to various real spider species.
‍
Species in the upper right corner are those the model finds most aligned with Bourgeois’s aesthetic taste, while those in the lower left are least aligned. Surprisingly, it showed a strong alignment with the Huntsman spider.

Sculpting with Loss Functions

04. Curating 3D Optimization

To translate aesthetic preferences into form, we used a volume-based 3D optimization script developed for this course. The process began with a starting image (in this case, a photo of a Huntsman spider) and a starting mesh, modeled after Bourgeois’s Cell series. We used eight different CAGE.OBJ models as the base geometry.

The optimization script reshapes this base form over thousands of iterations, guided by a custom-defined loss function. This loss encodes what the system should prefer. In our case, the key components were:
· SimilarityToImageLoss – encouraging outputs to stay close to the spider reference image
· ClassifierLoss – applying our Bourgeois taste classifier to push the generated views toward her aesthetic

Finally, we applied random affine transformations and Gaussian smoothing during training, regularizing the optimization process to avoid overfitting or pixel-specific artifacts.

The result: volumetric forms that emerge as emotionally weighted, almost architectural spider-sculptures, computationally shaped by Bourgeois’s aesthetic logic.

Encountering Form, Perception, and Machine Taste

05. Virtual Exhibition

The final spider sculptures are presented in a Unity-based virtual exhibition, placing machine-generated forms into an immersive spatial context.

At the center of the space is a piece created through creative coding:

From afar, it appears to be a cage, geometric, cold, and closed. But as the viewer approaches, the form subtly shifts, revealing itself to be a spider, generated through the AI optimization process.

Coding test in Unity: A distance-based morphing structure that shifts from one object into another as the viewer approaches.