🤖 AI

Probabilistic Dating of Historical Manuscripts via Evidential Deep Regression on Visual Script Features

This paper introduces a probabilistic deep learning framework that treats historical manuscript dating as a continuous regression problem, achieving state-of-the-art accuracy and superior uncertainty calibration on the DIVA-HisDB benchmark by outputting full predictive distributions with decomposed aleatoric and epistemic uncertainties in a single forward pass.

Original authors: Ranjith Chodavarapu

Published 2026-05-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ranjith Chodavarapu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a pile of ancient, handwritten letters from the Middle Ages. You want to know exactly when each one was written. Usually, experts (called paleographers) have to squint at the handwriting, guess the date, and argue about it for hours. It's slow, expensive, and even experts often disagree.

This paper introduces a new kind of "AI time machine" that doesn't just guess a date; it tells you how sure it is about that guess.

Here is the breakdown of how it works, using simple analogies:

1. The Old Way vs. The New Way

The Old Way (Classification): Imagine trying to sort mail into three big bins: "1800s," "1900s," and "2000s." The computer looks at the handwriting and says, "This is definitely from the 1900s." It's a rough guess. The paper says this is too vague.
The Point Guess (Regression): Imagine the computer giving you a specific year, like "1245." It's more precise, but it's like a dart thrower who hits the bullseye but won't tell you if they were lucky or if they are a master. If the computer is wrong, you have no idea.
The New Way (Probabilistic/Evidential): This is the paper's invention. Instead of throwing one dart, the computer throws a cloud of darts and draws a shape around where they landed. It says, "I think it's 1245, but I'm 90% sure it's between 1240 and 1250." Crucially, it breaks down why it's unsure:
- Aleatoric Uncertainty (The "Messy Ink" factor): The handwriting itself is faded, smudged, or written by a shaky hand. Even a human expert would be confused. The AI admits, "The evidence is just bad."
- Epistemic Uncertainty (The "I've Never Seen This" factor): The AI hasn't seen this style of writing before. It's saying, "I don't know what this is because I haven't studied it."

2. How the AI "Thinks"

The researchers built a neural network (a type of AI brain) that looks at tiny square patches of the manuscript pages.

The Brain: They used a pre-trained brain called EfficientNet, which is already good at recognizing shapes and textures.
The Special Head: They added a special "calculator" on top of the brain. Instead of just outputting a number (the year), it outputs a whole probability distribution. Think of it like a weather forecast that doesn't just say "It will rain," but gives you a bell curve showing the chance of light drizzle vs. a hurricane.
The Training: They taught the AI using a special math trick (Evidential Deep Regression) that forces it to learn both the date and its own confidence level in a single glance.

3. The Results: How Good Is It?

The team tested this on a famous dataset of 150 medieval pages (DIVA-HisDB).

Accuracy: The AI guessed the year with an average error of only 5.4 years. That is incredibly precise considering the handwriting is 1,000 years old.
Calibration (The "Trust" Meter): This is the paper's biggest win.
- Other methods (like "Deep Ensembles" or "MC Dropout") were either very accurate but overconfident (saying "I'm 100% sure" when they were wrong) or well-calibrated but very slow (needing 50 tries to get one answer).
- The new AI is fast (one look) and honest. When it says "I'm 90% confident," it is actually right 92.6% of the time. It's like a weatherman who is right almost every time they say "90% chance of rain."
The "Selective" Superpower: Because the AI knows when it is unsure, you can tell it to "only show me the answers you are 100% sure about."
- If you filter out the messy, uncertain pages and only look at the top 20% of confident predictions, the error drops to 0.5 years.
- Analogy: Imagine a librarian who says, "I can't read this smudged page, but for the 20 clearest pages, I can tell you the exact year to the month."

4. What the AI Sees

The researchers used "heat maps" (like thermal cameras) to see what the AI was looking at.

Good Clues: The AI focused on the actual ink strokes and letter shapes.
Bad Clues: When the ink was faded or the paper was blurry, the AI's "uncertainty meter" went up. It correctly identified that the image quality was the problem, not its own intelligence.
The "Faded Ink" Test: The AI found that when the ink was faint, its "Aleatoric uncertainty" (the "messy ink" factor) spiked. This proves the AI is actually learning about the physical state of the manuscript, not just memorizing patterns.

5. Why This Matters for History

The paper argues that this tool solves a major bottleneck in libraries.

The Workflow: Imagine a massive digitization project with millions of pages. You can't have a human expert look at every single one.
The Solution: Run the AI on everything.
- If the AI is confident, it automatically dates the page and files it away.
- If the AI is uncertain (because the ink is faded or the handwriting is weird), it flags that page and sends it to a human expert for a second look.
The Benefit: This saves time and money. It lets humans focus only on the difficult cases, while the AI handles the easy, clear ones.

Summary

This paper presents a new AI that doesn't just guess the age of ancient handwriting; it acts like a humble expert who knows the limits of its own knowledge. It is fast, accurate, and, most importantly, it tells you when it is guessing and when it is certain, allowing historians to trust the machine for the easy jobs and step in for the hard ones.

Technical Summary: Probabilistic Dating of Historical Manuscripts via Evidential Deep Regression

Problem Formulation
The paper addresses the challenge of estimating the authorship date of historical manuscripts using only visual features. Traditional automated approaches have relied on either classification (discretizing dates into century-sized buckets) or point regression (predicting a single year). The authors argue that classification lacks precision (resulting in high Mean Absolute Error, or MAE), while point regression fails to quantify confidence, leaving historians without the necessary signal to assess reliability. The core problem is reframed as a probabilistic inference task: instead of a single point estimate, the goal is to output a full predictive distribution over a continuous year axis that explicitly decomposes uncertainty into aleatoric (irreducible data noise, such as scribal ambiguity) and epistemic (model ignorance) components.

Methodology
The proposed solution utilizes Evidential Deep Regression, a framework that models the parameters of a likelihood function using a higher-order prior.

Architecture: The model employs an EfficientNet-B2 backbone (pretrained on ImageNet) to extract visual features from manuscript patches (224×224 pixels). These features are passed through a fully connected layer and a Normal-Inverse-Gamma (NIG) output head.
Output: The network outputs four parameters ( $\gamma, \nu, \alpha, \beta$ $γ, ν, α, β$ ) defining a NIG distribution. The predictive mean is $\hat{y} = \gamma$ $\overset{y}{^} = γ$ . The total predictive variance is analytically decomposed into:
- Aleatoric uncertainty: $\frac{\beta}{\alpha-1}$ , representing irreducible noise in the data (e.g., faded ink, mixed hands).
- Epistemic uncertainty: Scaled by $\frac{1}{\nu}$ , representing model ignorance, which decreases as the model accumulates evidence.
Training Objective: The model is trained using a joint loss function comprising the Negative Log-Likelihood (NLL) of the NIG distribution and an evidence-regularization term. This regularizer penalizes the accumulation of high evidence ( $\nu$ ) for incorrect predictions, preventing the model from becoming overconfident on erroneous data.
Inference: Unlike Monte Carlo Dropout (which requires 50 stochastic passes) or Deep Ensembles (which require 5 models), this approach generates the full predictive distribution and decomposed uncertainties in a single forward pass.

Key Contributions

First Probabilistic Formulation: This work introduces the first evidential deep regression framework for manuscript dating, enabling the decomposition of uncertainty into aleatoric and epistemic components.
Calibration and Efficiency: The model achieves the best calibration among compared uncertainty quantification (UQ) methods (PICP=92.6%) while maintaining competitive accuracy, all within a single inference pass.
Selective Prediction: The framework enables "selective prediction," where the system can filter out uncertain predictions. By retaining only the top 20% most certain patches, the model achieves sub-year accuracy (0.5 years MAE) despite being trained with century-level supervision granularity.
Uncertainty-Error Correlation: The study demonstrates that aleatoric uncertainty is a strong predictor of dating error (Spearman $\rho = 0.729$ ), allowing the model to identify specific script regions (e.g., faint ink, ligatures) that cause ambiguity.

Experimental Results
Evaluated on the DIVA-HisDB benchmark (150 pages from three medieval codices, 151,936 patches), the model achieved the following:

Accuracy: A test MAE of 5.4 years, significantly outperforming a CNN classifier baseline (47.1 years) and comparable to a point regressor (2.8 years) but with the added benefit of uncertainty estimation.
Precision: 93% of patches were dated within 5 years, and 97% within 10 years.
Calibration: The model achieved a Prediction Interval Coverage Probability (PICP) of 92.6% at the 90% confidence level. This outperformed MC Dropout (88.2% PICP, requiring 50 passes) and Deep Ensembles (79.7% PICP, requiring 5 models).
Degradation Robustness: Predicted uncertainty increased monotonically with image degradation (blur and JPEG compression), serving as a practical indicator for image quality and the need for re-scanning.
Page-Level Aggregation: Aggregating patch predictions to the page level reduced MAE to 4.5 years and established a near-perfect correlation ( $\rho = 0.905$ ) between page-level uncertainty and page-level error, effectively ranking pages by dating difficulty.

Significance and Claims
The paper claims that this approach directly addresses a bottleneck in the mass digitization of medieval manuscripts. By providing a calibrated probabilistic estimate, the system enables a workflow where certain predictions are automated while uncertain ones are flagged for expert review. The authors emphasize that the ability to trade coverage for accuracy (via uncertainty thresholds) is a unique capability of probabilistic models, offering a practical tool for curators and historians.

However, the paper maintains a modest scope regarding limitations. It acknowledges that the model currently acts as an interpolator within the known distribution of the three training script families. It notes that epistemic uncertainty did not spike for out-of-distribution (OOD) manuscripts, suggesting the model lacks the "stylistic overlap" needed to recognize truly novel scripts as unfamiliar. The authors suggest that future deployments would require more diverse training data (e.g., CLaMM or e-codices) to improve generalization to unseen script families.