Computer Vision Meets the Harappan Seals: What Hybrid Embeddings Can—and Cannot—Tell Us About Indus Script Similarity

Archaeologists who spend their seasons in the trench know that visual surface patterns can be deceiving. A rim sherd from a late occupation layer can look identical to a vessel rim from five centuries earlier until you trace the stratigraphy, analyze the temper under a lens, and map its precise micro-context.

Epigraphy faces a similar challenge when computational tools are applied to undeciphered scripts. In a paper led by Ooha Lakkadi Reddy (arXiv, 2025), researchers applied a hybrid convolutional-transformer neural network architecture to evaluate visual similarity between signs of the Indus script—found on steatite (soapstone) seals, terracotta tablets, and copper plates across the Harappan sphere—and the pictographic writing systems of the Tibetan–Yi Corridor (arXiv, 2025).

Using an ensemble of 15 independently trained vision models, the study calculated vector embeddings for individual sign images (Zenodo, 2026). The authors report that glyphs from the Tibetan–Yi Corridor displayed a mean cosine similarity score of 0.635 when compared to Indus script signs (Zenodo, 2026). By contrast, comparison of the Indus signs against West Asian Bronze Age scripts yielded lower visual similarity scores: 0.102 for Proto-Cuneiform and 0.078 for Proto-Elamite (Zenodo, 2026). In direct script-to-script embedding space, the model mapped the Indus signary closest to Tibetan–Yi systems with a mean similarity of 0.930 (Zenodo, 2026).

The statistical separation in vector space is clear. What remains debated is how to interpret vector distances when stripped of archaeological and linguistic context.

What Embeddings Actually Measure

Computer vision embeddings work by breaking down pixel representations of shapes into mathematical vectors. Convolutional layers extract local geometric strokes—loops, intersecting lines, parallel hatches, and vertical stems. Transformer layers then evaluate how these low-level features relate spatially across the image frame.

When an embedding algorithm reports high cosine similarity between two signs, it indicates that the pixel distributions share visual geometry. It does not measure linguistic syntax, underlying phonology, or semantic meaning.

Human stroke construction poses an immediate constraint on shape variation. The human hand drawing a pictograph on clay or carving a glyph into soft stone operates under mechanical limits. Vertical lines, simple enclosures, hatch marks, and animal outlines recur across completely unrelated writing systems across the globe.

Evidence shows that structural similarity in visual space often arises from convergent evolution rather than historical transmission. Without a bilingual text, visual similarity alone cannot establish a genetic or historical link between two sign systems.

The Problem of Decontextualized Glyphs

Traditional epigraphic analysis—as recorded in major corpora by Iravatham Mahadevan (The Indus Script: Texts, Concordance and Tables, Memoirs of the Archaeological Survey of India, 1977) and Asko Parpola—relies on positional syntax and contextual association. Epigraphers analyze how signs cluster in sequence, which signs appear exclusively in initial or final positions, and how glyph variations correlate with artifact types.

A computer vision model processing isolated sign images risks ignoring three critical factors:

1. Medium Constraints: An Indus sign carved into a miniature steatite seal measuring two centimeters square encounters different physical constraints than a symbol painted on wood or inked on paper. Surface texture, tool width, and weathering alter stroke thickness.
2. Positional Context: In the Indus corpus, specific signs occur almost exclusively at the end of sequences. A visual embedding model comparing individual glyphs in isolation treats a terminal sign identically to a medially positioned sign, disregarding positional rules.
3. Chronological Layering: The Indus script was used over several centuries, spanning the Urban Mature Harappan phase (c. 2600–1900 BCE). Sign forms underwent subtle modifications over time. Aggregating all sign forms into a single visual corpus flattens chronological developments into a static dataset.

Historical and Spatial Constraints

The paper suggests potential contact along historical trade routes, such as the ancient Shu–Shendu passage, as a theoretical explanation for the observed visual overlap (arXiv, 2025). However, testing such hypotheses requires corroboration from physical stratigraphy and material culture.

Evidence shows extensive Bronze Age trade networks linking the Indus Valley to the Persian Gulf, Oman, and Mesopotamia. This trade is documented by Indus carnelian beads, square steatite seals, and standardized cubical chert weights recovered from stratified contexts in West Asia.

By contrast, material evidence for direct, sustained trade between the Harappan core and the high-altitude Tibetan–Yi Corridor during the mid-third millennium BCE remains sparse in the archaeological record published in journals such as Antiquity, Puratattva, and Man and Environment. While West Asian scripts share lower visual similarity scores in the computer vision embedding space, historical trade contacts between the Indus and Mesopotamia are firmly anchored by physical artifacts recovered from excavated strata.

⚖ Scholarly debate: Visual Clustering vs. Epigraphic Decipherment

Scholars debate the role that machine learning should play in analyzing undeciphered writing systems.

One perspective holds that automated vision models provide an objective, quantitative tool for mapping sign variants across large digital corpora. By clustering signs according to objective mathematical parameters, machine learning can identify overlooked graphic variants and assist in standardizing sign catalogues without human bias.

The opposing view argues that visual similarity models risk generating false positives by matching superficially similar geometric shapes across unrelated cultures. Because script functions as a visual representation of language or systematic information, surface alignment alone cannot establish structural or linguistic relationships. Epigraphers emphasize that internal positional statistics, positional combinatorics, and stratigraphically verified site distribution must take precedence over visual clustering algorithms.

Computational tools offer powerful methods for organizing digital catalogues and accelerating image segmentation. However, until an archaeological discovery provides a bilingual key or a stratigraphically grounded sequence, mathematical distance vectors can map visual shapes, but they cannot decipher their meaning.