A Large-Scale Computer-Vision Mapping of the Geometric Structures of Stroboscopically-Induced Visual Hallucinations

1. Problem Statement

Visual hallucinations (VHs) are prevalent across various conditions (psychiatric, neurological, and psychedelic states), yet their phenomenology remains difficult to characterize systematically. Empirical research is hindered by a lack of large-scale, quantitative datasets.

• The Gap: Existing theoretical frameworks, specifically Klüver's form constants (lattices, cobwebs, tunnels, spirals) and neural-field models (based on retinocortical log-polar transforms), propose a limited set of geometric patterns. However, these models are weakly supported by large-scale empirical data and fail to account for geometric structures outside their predicted symmetry groups (e.g., translational vs. rotational symmetry).
• The Challenge: Traditional methods rely on small samples, heterogeneous report formats, and subjective descriptions, making it difficult to determine the true diversity of simple VH geometries or to rigorously test mechanistic models.

2. Methodology

The authors employed a mixed-methods approach combining a massive, naturally occurring dataset with unsupervised computer vision techniques.

A. Data Collection

• Source: 10,598 drawings created by attendees of the Dreamachine public art-science installation (UK, 2022).
• Stimulus: The installation used Stroboscopic Light Stimulation (SLS) (High Sensory format) to induce visual hallucinations (SIVHs) in closed-eyed participants. A minority of attendees experienced a non-stroboscopic "Deep Listening" format (serving as a noise control).
• Procedure: Participants drew their experiences using colored pastels on A3 paper. The dataset was fully anonymized and publicly released by the organizers.

B. Data Preprocessing

• Filtering: The initial 10,598 images were cleaned to remove duplicates, corrupted files, text-only reflections, and improperly cropped scans.
• Final Dataset: 9,700 usable drawings.
• Normalization: Images were center-cropped, resized to 448x448 pixels, and subjected to histogram equalization and Gaussian smoothing to reduce lighting artifacts.

C. Feature Extraction (Computer Vision Pipeline)

• Model: A pretrained DINOv2 Vision Transformer (ViT-L/14), a self-supervised model trained on 142 million natural images, was used to extract feature embeddings.
• Test-Time Augmentation (TTA): To ensure robustness against orientation and color variations, 48 augmented views (rotations, flips, color jitter) were generated for each drawing. The resulting 48 feature vectors (1,024 dimensions each) were averaged to produce a single robust embedding per drawing.

D. Dimensionality Reduction and Clustering

• Reduction: Features were first reduced via Principal Component Analysis (PCA) (retaining 90% variance) and then projected into a 6-dimensional space using UMAP (Uniform Manifold Approximation and Projection).
• Clustering: HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise) was applied to the 6D embeddings. This algorithm was chosen for its ability to handle clusters of varying density and to label outliers as noise rather than forcing them into spherical clusters.
• Validation: The solution was optimized using the Density Based Clustering Validity (DBCV) metric and qualitative interpretability checks.

E. Qualitative Analysis

• Two authors performed an inductive thematic analysis on the resulting clusters, assigning descriptive codes and mapping them to three higher-level categories: Klüverian (matching classical form constants), Non-Klüverian (recurring but unreported geometries), and Non-Geometric.

3. Key Contributions

1. Scale: The study analyzes the largest dataset of VH drawings to date (n=9,700), moving beyond small-scale case studies.
2. Methodological Innovation: It demonstrates the efficacy of an unsupervised computer-vision pipeline (ViT + HDBSCAN) for organizing large-scale phenomenological data without relying on pre-defined categories.
3. Empirical Expansion: It identifies a significantly broader range of geometric motifs than previously documented, challenging the completeness of existing taxonomies.
4. Theoretical Constraints: The findings provide new empirical constraints for neural-field models of visual hallucinations, specifically regarding symmetry types.

4. Results

The HDBSCAN algorithm successfully clustered 59.1% (5,736) of the drawings into 15 distinct clusters, while 40.9% were classified as noise (idiosyncratic or mixed content).

A. Thematic Categories

1. Klüverian Geometries (40.9% of total drawings):
   • Contained patterns matching the four classical form constants: Tunnels, Spirals, Cobwebs, and Lattices.
   • These were predominantly rotationally symmetric (tunnels, spirals) or translationally symmetric (lattices).
2. Non-Klüverian Geometries (9.9% of total drawings):
   • Identified systematically recurring patterns not captured by Klüver's framework.
   • Key Discoveries:
     • Fourfold Rotational Symmetry: Concentric squares, crosses, and "diamond lattices" (Cluster 13 & 14).
     • Astroid Patterns: Rounded, pinched-square curves.
     • Wavy/Zig-zag Stripes: Undulating geometries.
   • These forms appeared consistently across all four UK venues, ruling out site-specific artifacts.
3. Non-Geometric Content (8.3%):
   • Included diffuse color fields, semantic imagery (faces, objects), and text/scribbles.
4. Unclustered (40.9%):
   • Represented highly idiosyncratic drawings or mixtures of features that did not meet the density threshold for clustering.

B. Statistical Findings

• There was a statistically significant but negligible association between cluster type and geographic location, confirming the robustness of the geometric motifs across different populations.
• The prevalence of fourfold rotational symmetry (approx. 5.8% of geometric drawings) suggests it is a distinct, under-characterized subclass of simple VHs.

5. Significance and Implications

• Challenging Neural-Field Models: Current models based on the log-polar retinocortical transform successfully predict rotationally symmetric patterns (spirals, tunnels) but struggle to explain translational symmetry (lattices) and fourfold rotational symmetry (squares, crosses). The data suggests these models are incomplete and require extension to account for the full spectrum of observed geometries.
• Mechanistic Insight: The recurrence of specific, structured geometries in a large, heterogeneous sample implies that simple VHs are driven by the intrinsic architecture and dynamics of the early visual cortex (V1), rather than random noise or top-down cognitive processes alone.
• Future Directions: The study establishes a framework for "quantitative phenomenology." It motivates targeted experimental work to link specific SLS parameters (frequency, brightness) to the emergence of these novel geometric classes, potentially refining our understanding of consciousness and visual processing.

In conclusion, this paper bridges the gap between large-scale participatory data and computational neuroscience, revealing that the "space" of simple visual hallucinations is far more diverse and structured than classical theories have acknowledged.