Quantifying the Occult: A Comparative Study of Hindu and Buddhist Deities Using Machine Learning Methods

1 Introduction

• The framework aims to quantify how esoteric traditions preserve core cosmic functions while altering physical forms across cultural settings.The introduction also proposes modeling how prominent orthodox symbols can draw attention from underlying Tantric forms.Source: §1 · p. 2
• The study introduces an open-source dataset of 196 cataloged Hindu and Vajrayana Buddhist deities, structured to evaluate visual and theological topologies.The dataset is presented as a granular alternative to standard literary corpora.Source: §1 · pp. 1–2
• A dual-matrix architecture measures morphological divergence with discrete Gower distance and theological semantics with continuous embeddings.The framework is proposed for modeling phenomena such as iconographic camouflage and the Atin Effect.Source: §1 · p. 2
• Cardinality Weighting emphasizes rare, high-variance esoteric symbols so generic aesthetic traits do not suppress their topological weight.The modification is presented as a response to the Gestalt Problem in visual contexts.Source: §1 · p. 2
• Dense word embeddings represent theological domains in continuous semantic space, enabling comparison of concepts such as Lakshmi’s Wealth and Vasudhara’s Abundance.The semantic vectors provide a functional baseline against which visual camouflage can be measured.Source: §1 · p. 2

2 Related Work

• The study positions its approach at the intersection of computational iconography, digital humanities, and Tantric studies, where such topological mapping is described as virtually unprecedented.The literature review is organized across three parallel scholarship domains.Source: §2 · pp. 2–3

2.1 Computational Art History and Iconography

• Prior CNN and deep-learning approaches classify or identify visual culture but require large labeled datasets and operate as black boxes for feature-sensitive theological analysis.The paper contrasts these methods with unsupervised topological mapping and Gower distance for bounded, high-complexity domains such as N = 196.Source: §2.1 · p. 3

2.2 Quantitative Mythography and Network Analysis

• Earlier mythography often maps deity interactions through textual co-occurrences, whereas this study compares intrinsic iconographic architecture using visual and semantic matrices.The cited methodological precedent treats cultural objects as relational data.Source: §2.2 · p. 3

2.3 Comparative Tantric Studies

• Comparative scholarship has documented Hindu–Buddhist borrowing and near-identical deity presentations, but those observations have remained qualitative; this study maps them computationally.The cited scholarship includes Sanderson and Benard.Source: §2.3 · p. 3

3 Methodology

• The dual-matrix framework measures visual “additive gravity” with Cardinality Weighting and theological context with semantic vectorization.Together, the matrices are intended to test visual camouflage and cognitive bias in esoteric iconography.Source: §3 · pp. 3–4

3.1 Data Collection and Feature Engineering

• The dataset contains 196 deities whose documented physical and theological parameters were cross-checked against canonical texts, ritual manuals, and scholarly editions.Visual features include anatomical counts, aesthetic categories, wrathful markers, and arrays of specific implements and mudras.Source: §3.1 · p. 4

3.2 Vector Space Modeling and the “Gestalt Problem”

• Standard logarithmic TF-IDF suppressed rare identity-defining symbols, producing a false-positive Mandarava–Lakshmi cluster from shared peaceful traits.The study calls this distortion the “Gestalt Problem”; the two deities’ esoteric implements diverge.Source: §3.2 · pp. 4–5

3.3 Cardinality Weighting: An Objective Heuristic for Esoteric Gravity

• Cardinality Weighting multiplies inverse-frequency feature weights by category cardinality, elevating rare objects such as the veena over common traits such as apparent age.The method derives feature importance from dataset variance rather than hardcoded human multipliers, and is intended to reduce the Mandarava–Lakshmi anomaly.Source: §3.3 · p. 5

3.4 Semantic Vectorization (The Theological Matrix)

• Dense word embeddings map deity domains and descriptions into continuous semantic space, allowing contextual proximity beyond exact text matches.The semantic matrix is framed as a functional baseline, but as a synthetic proxy rather than definitive historical ground truth.Source: §3.4 · pp. 5–6

3.5 Similarity Metrics and Additive Gravity

• The visual matrix uses an unnormalized dot product so shared rare symbols add linearly to similarity, then inverts and normalizes the score into a 0.0–1.0 distance.The resulting pairwise matrix covers all 196 deities and supports comparative ranking and spatial visualization.Source: §3.5 · p. 6

4 Implementation Details

• The architecture models 196 Hindu and Vajrayana deities in independent visual-morphology and theological-function modalities.The visual feature vector contains 28 mixed-type variables, including categorical, binary, and ordinal attributes.Source: §4 · p. 7, §4.1 · p. 7

4.1 The Visual Matrix: Morphological Encoding and Gower Distance

• Gower distance measures pairwise morphological divergence across mixed feature types, using feature weights and observable-value indicators.The distance is calculated as one minus Gower similarity.Source: §4.1 · p. 7

4.2 The Semantic Matrix: Generative Embeddings and Vector Space

• The semantic pipeline expands each core cosmic domain with an LLM, then embeds the text as a 768-dimensional vector for cosine-distance comparison.The expansion model is Gemini 3.1 Flash Lite; the sentence-transformer is all-mpnet-base-v2.Source: §4.2 · p. 7

4.3 Topological Projection and Normalization

• UMAP projects the precomputed visual and semantic distance matrices, and target-relative Min-Max normalization synchronizes radial divergence to a 0.0–1.0 coefficient.UMAP uses n neighbors = 12 and min dist = 0.1; scripts and visualizations were implemented in Python 3.12.Source: §4.3 · p. 8

5.1 Visual Matrix Proximities: Nearest Neighbours of Selected Deities

• Kali’s visual-neighbor matrix forms a localized cluster of dark, wrathful deities bearing the khadga and severed head.The selected visual-neighbor analyses examine targets chosen to illustrate anchor dependency, the Atin Effect, and form-versus-function parallax.Source: §5.1 · p. 8, Table 1 · p. 8
• Marichi shares four nearest neighbors with Durga at D = 0.3416, while Vasudhara is visually pulled toward Dhanya Lakshmi by the shared grain-sheaf anchor.The paper presents the warrior pairing as structural mirroring and the Vasudhara result as visual camouflage despite theological correspondence with Lakshmi.Source: §5.1 · p. 9, Table 7 · p. 9, Table 8 · p. 9

5.2 Semantic Matrix Proximities: Theologically Nearest Neighbours

• The semantic matrix links Lakshmi with Vasudhara through wealth and abundance, and bridges Saraswati and Matangi to Hayagriva Dakini through speech, mantras, and primordial sound.Marichi’s semantic network collapses to uniform D = 0.0000 distances; the paper attributes this limitation to specialized out-of-vocabulary terms producing null vectors.Source: §5.2 · pp. 10–11

5.3 Global Topological Mapping: UMAP Projection

• In the 3D projection, clusters follow the gravitational pull of esoteric anchors rather than Hindu–Buddhist tradition, with wrathful implements drawing entities away from the peaceful orthodox manifold.The projection is presented as a dimensionality-reduced view of the visual matrix.Source: §5.3 · p. 11, Figure 1 · p. 12

5.4 Micro-Gravitational Orbits: Dual Egocentric Networks

• Lakshmi and Vasudhara have virtually no visual overlap but converge in the inner semantic orbit through shared wealth, abundance, and preservation domains.Krodhakali remains maximally divergent (1.0) across both matrices.Source: §5.4 · p. 12, Figure 2 · p. 13

5.5 The Atin Effect: Sequential Perception and Orthogonal Anchoring

• The Atin Effect describes how encountering Matangi’s veena first can trigger a Saraswati schema, while the model evaluates the veena alongside her wrathful implements.The paper characterizes human perception as sequential and the model’s feature evaluation as parallel.Source: §5.5 · pp. 13–14

5.6 Iconographic Camouflage and Boundary Mapping

• Cosine Distance = 0.068 and Gower Distance = 0.288 place Chinnamasta and Chinnamunda near-identically across semantic and visual matrices, respectively.The paper describes this as a 1:1 esoteric transfer across religious boundaries.Source: §5.6 · p. 14, Table 15 · p. 15

Generative Inference and Theological Resolution

• LLM-generated theological expansions remain unvalidated against primary Sanskrit or Tibetan corpora, and their agreement with secondary sources is not independent validation.The authors identify access to digitized primary texts and Sanskrit and Tibetan philological expertise as requirements for ground-truth validation.Source: §6 · p. 15

Discrete Morphology and the Loss of Art Style

• The discrete visual matrix omits artistic style, medium, and historical era, placing matching categorical attributes at the same coordinate.The authors contrast a 10th-century Chola bronze and a 19th-century Kalighat lithograph of the same deity.Source: §6 · p. 15

Bounded Dataset and Topological Scope

• The N = 196 dataset captures primary Hindu, Dashamahavidya, and Vajrayana nodes but underrepresents highly localized folk Tantric traditions.The authors identify dataset expansion as a future objective.Source: §6 · pp. 15–16

7 Conclusion

• The dual-matrix mapping identifies iconographic camouflage and places Chinnamasta and Chinnamunda at near-identical topological coordinates across sectarian boundaries.The authors describe the mapping as mathematical validation of historical qualitative scholarship and as tracing a shared continuum of cosmic functions.Source: §7 · p. 16
• The authors present the architecture as a scalable, reproducible framework for mapping continuums across historical, literary, or anthropological corpora.It treats qualitative descriptors as discrete anatomical anchors and uses semantic proxies for functional domains.Source: §7 · p. 16
• Standardized iconographic anchors are interpreted as compact carriers of complex psychological and cosmic domains across generations.The paper likens these anchors to an informational architecture for transmitting esoteric technologies.Source: §7 · p. 16
• Multidimensional scaling cannot capture the lived, unquantifiable resonance of deities within active practitioner lineages.This qualifies the computational lens as a statistical approach to comparative iconography.Source: §7 · p. 16

Data and Code Availability

• The supplementary repository includes the 196-node morphological dataset, code for distance and projection calculations, an egocentric-network script, the semantic distance matrix, and interactive plots.The materials are presented as supporting transparency and computational reproducibility.Source: Data and Code Availability · pp. 16–17

Links

• The study links a complete code repository on GitHub and a public dataset on Kaggle.The links are listed under the paper’s Links subsection.Source: Links · p. 17
• Researchers are encouraged to verify cross-tradition bridges, test the Atin Effect on out-of-sample entities, and expand the matrices to additional pantheons.Source: Links · p. 17

A.1 System Prompt

• The appendix reproduces the Gemini 3.1 Flash Lite system prompt to document the generative-inference pipeline.The prompt is described as using formatting constraints and few-shot examples to limit narrative hallucination and elicit dense ontological fields.Source: A LLM System Prompt and Semantic… · p. 19, §A.1 · p. 19
• The prompt assigns an expert role and requests 15–25 nouns or noun phrases representing a deity’s esoteric semantic field.Source: §A.1 · p. 19
• Its rules prioritize Sanskrit equivalents and functionally adjacent concepts, exclude appearance descriptors, favor specific terms, and require comma-separated output only.Source: §A.1 · p. 19

A.2 In-Context Examples and Generated Nouns

• Few-shot examples show generated noun fields for Nila Saraswati, Matangi, Kali, Saraswati, and Chinnamasta.The examples include sound, speech, knowledge, dissolution, and self-offering domains.Source: §A.2 · pp. 19–20

A.3 Dynamic Semantic Generation for Novel Domains

• For unseen domains, the appendix presents outputs for Bagalamukhi, Bhuvaneshwari, Bhairavi, Tara, and Kurukulla, with target domains drawn from textual traditions.Source: §A.3 · pp. 20–21
• Additional outputs cover Dhumavati, Pratyangira, Tripura Sundari, Vajrayogini, and Chakrasamvara.Their stated domains include inauspiciousness and void, black-magic reversal, beauty and sovereignty, bliss and enlightenment, and supreme bliss.Source: §A.3 · pp. 20–21

• The section records the architecture’s morphological variables, mathematical hyperparameters, and environmental dependencies for transparency and reproducibility.Source: B Methodological and Computational… · p. 21

• The Visual Matrix uses a 28-dimensional discrete Gower distance that accommodates mixed data types and encodes nominal features as exact matches or mismatches.Identical strings receive distance 0 and mismatches receive distance 1, allowing objects such as a veena or shuka to be encoded.Source: §B.1 · p. 21
• The morphological dataset columns are categorized by data type, and the dataset is publicly available on Kaggle and GitHub.The links are given in the Data & Code Availability section.Source: §B.1 · p. 21

B.2 Computational Hyperparameters

• UMAP settings include a minimum distance of 0.1, a precomputed metric, random state 42, and 15 neighbors; changing parameters can warp spatial topologies.The stated minimum-distance rationale is to allow tight clustering of functionally identical esoteric entities.Source: §B.2 · pp. 21–22
• The 28 morphological variables are the mixed-type inputs to the discrete Gower Distance matrix.Source: Table 16 · p. 22
• Generative semantic inference uses gemini-3.1-flash-lite with strict formatting and in-context few-shot prompting, while dense embeddings use all-mpnet-base-v2 and pairwise cosine distance.Source: §B.2 · p. 22

B.3 Software Environment

• The processing scripts ran in a standard Python 3 Kaggle kernel using pandas, numpy, scipy, umap-learn, sentence-transformers, and google-generativeai.The codebase is linked in the Data and Code Availability section.Source: §B.3 · p. 22

B.4 Hardware and Computational Execution

• An NVIDIA Tesla T4 GPU was used for semantic inference, 768-dimensional theological embeddings, and UMAP projections.The passage also reports computation of the discrete R28 morphological distance matrices.Source: §B.4 · p. 22