💻 computer science

Disentangling Model and Human Data Uncertainty in Apparent Facial Age Estimation

This paper demonstrates that Bayesian Neural Networks trained on the APPA-REAL dataset can effectively disentangle and quantify aleatoric and epistemic uncertainties in facial age estimation, revealing that aleatoric uncertainty remains stable across dataset sizes while epistemic uncertainty increases as training data decreases.

Original authors: Andrei Foitos, Ivo Pascal de Jong, Matias Valdenegro-Toro

Published 2026-07-21

📖 4 min read☕ Coffee break read

CC0 1.0

Original authors: Andrei Foitos, Ivo Pascal de Jong, Matias Valdenegro-Toro

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ 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 are trying to guess how old a stranger is just by looking at a photo. You might say, "They look about 30," but your friend might say, "No way, they look 40!" In the world of artificial intelligence, this isn't just a harmless disagreement; it's a massive puzzle. AI systems are getting better at looking at faces and guessing ages, but they often act like overconfident know-it-alls. They give you a single number and act like it's the absolute truth, even when the answer is fuzzy. This is dangerous because if a computer is too sure of a wrong answer, it could make bad decisions, like letting a teenager buy alcohol or blocking a senior citizen from a website. To fix this, scientists are teaching AI to admit when it's guessing. They want the computer to say, "I think they are 30, but I'm only 50% sure," or "I think they are 30, but I'm very sure." To do this, researchers split "uncertainty" into two flavors: aleatoric uncertainty, which is the natural messiness of the data (like how humans just can't agree on an age), and epistemic uncertainty, which is the AI's own ignorance (like a student who hasn't studied enough).

This paper is like a detective story where the investigators try to teach a computer to tell the difference between "the data is confusing" and "I am confused." The researchers used a special dataset called APPA-REAL, where thousands of real humans looked at the same faces and voted on how old they looked. Because so many people voted, the researchers could measure exactly how much humans disagreed (the "messiness") and use that to train the AI. They built three different types of "Bayesian" AI models—think of these as three different teams of detectives, each using a slightly different strategy to guess and measure their own confidence. They tested these teams by feeding them different amounts of data, from a tiny handful of photos to the entire library.

The big discovery here is that the two types of uncertainty behave very differently, and the AI can learn to separate them. The "messiness" of the data (aleatoric uncertainty) stayed exactly the same no matter how many photos the AI saw. It's like a noisy room: whether you have one person or a thousand people listening, the background noise level doesn't change. The paper found this noise level was stable at about 18 years squared (which translates to a standard deviation of roughly 4.2 years). This confirms that some confusion is just built into the nature of human perception.

On the other hand, the AI's own ignorance (epistemic uncertainty) acted exactly as expected: it dropped dramatically as the AI saw more data. When the AI only saw 1% of the training data (about 40 images), it was wildly unsure. But as they fed it more data, up to 100% of the set, the AI became much more confident. For example, the "DropConnect" method saw its epistemic uncertainty drop from about 12.9 years squared at 5% data down to 6.0 years squared with the full dataset. The "Deep Ensembles" method was the most accurate overall, finishing with a Mean Absolute Error (MAE) of 8.0 years, while "Flipout" was a bit slower to learn, ending with an MAE of 8.9 years.

The paper also tested if these smart AI detectives could handle a new, different set of faces (the UTKFace dataset). The results showed that the AI's confidence levels shifted appropriately when the data changed, proving the models could generalize their understanding of uncertainty. However, the authors note that the "Flipout" method was a bit wobbly and unstable, especially when data was scarce, suggesting it might not be the best choice for situations where the AI needs to be very reliable with limited information.

Ultimately, this research suggests that separating these two types of uncertainty is not only possible but actually works better in age estimation (a regression task) than in simple yes/no classification tasks. The study concludes that while Deep Ensembles are the most accurate, the DropConnect method offers the clearest picture of how the AI's confidence grows with more data. This means we are one step closer to AI that doesn't just guess your age, but also knows when it's just guessing.

Technical Summary: Disentangling Model and Human Data Uncertainty in Apparent Facial Age Estimation

Problem Statement
Facial age estimation is a challenging regression task due to the subjective nature of "apparent age," which relies on visual aging features rather than chronological reality. Unlike standard regression where a single ground truth exists, apparent age estimation involves inherent human disagreement; different annotators may assign different ages to the same image based on individual bias or cultural perception. Consequently, a robust model must not only predict the most likely age but also quantify the reliability of that prediction. Current machine learning models often output a single point estimate, failing to distinguish between two distinct sources of uncertainty:

Aleatoric Uncertainty: Inherent noise in the data (e.g., annotator disagreement).
Epistemic Uncertainty: Uncertainty arising from the model's lack of knowledge (e.g., limited training data, poor generalization).

While methods to disentangle these uncertainties exist, recent literature suggests they often fail to separate them effectively in classification tasks. It remains unclear if this disentanglement is successful in regression tasks like age estimation.

Methodology
The authors utilize the APPA-REAL dataset, which contains 7,591 images annotated by multiple human raters, providing both a mean apparent age and the standard deviation of those ratings. This dataset allows for the direct supervision of human aleatoric uncertainty.

Model Architecture: The proposed system is a multi-output regression architecture based on a DenseNet-121 backbone (pre-trained on ImageNet). The network branches into two heads:
- A Mean Branch predicting the apparent age.
- A Variance Branch predicting the standard deviation (aleatoric uncertainty), using a Softplus activation to ensure positive values.
Training Strategy: The model is trained using a three-phase pipeline: initial single-task training (mean only), multitask training (mean and variance), and fine-tuning. The variance branch is explicitly supervised using the standard deviation of human votes from the dataset, treating human disagreement as the ground truth for aleatoric uncertainty.
Uncertainty Disentanglement: To separate aleatoric and epistemic uncertainty, the authors employ three Bayesian Neural Network (BNN) approximations:
1. MC-DropConnect
2. Flipout
3. Deep Ensembles
  Epistemic uncertainty is calculated as the variance of predictions across multiple forward passes (or ensemble members) for the same input, while aleatoric uncertainty is the learned variance of the output distribution.
Experimental Design: Models were trained on seven different subset sizes of the dataset (1%, 5%, 10%, 25%, 50%, 75%, 100%) to observe how uncertainty metrics scale with data availability.

Key Contributions

Explicit Supervision of Human Uncertainty: The study incorporates human annotator disagreement directly into the training process, allowing the model to learn aleatoric uncertainty as a specific target rather than inferring it indirectly.
Disentanglement in Regression: The paper investigates whether aleatoric and epistemic uncertainties can be successfully separated in a regression setting, contrasting with previous findings in classification.
Empirical Analysis of Data Scaling: The authors systematically evaluate how different uncertainty estimation methods behave as training data volume changes, specifically testing the hypothesis that epistemic uncertainty decreases with more data while aleatoric uncertainty remains stable.

Results

Uncertainty Trends:
- Aleatoric Uncertainty: Remained largely stable across all dataset sizes (stabilizing at approximately 18 years², or ~4.2 years standard deviation). This confirms that human disagreement is an inherent property of the data, independent of the model's training size.
- Epistemic Uncertainty: Exhibited an inverse relationship with dataset size. As training data increased, epistemic uncertainty decreased. DropConnect showed the most pronounced reduction (from ~12.9 years² at 5% data to ~6.0 years² at 100%), while Deep Ensembles maintained consistently low epistemic uncertainty regardless of data size. Flipout showed less stable, non-monotonic behavior.
Predictive Performance:
- Deep Ensembles achieved the best overall predictive accuracy (Mean Absolute Error of 8.0 years at 100% data, $R^2$ of 0.61).
- DropConnect followed closely (MAE 8.3 years, $R^2$ 0.58).
- Flipout converged more slowly (MAE 8.9 years, $R^2$ 0.53).
Generalization: When tested on the UTKFace dataset, models showed increased error rates. However, the uncertainty metrics generally reflected this degradation, with epistemic uncertainty increasing for DropConnect and Ensembles, suggesting the models can detect distribution shifts.

Significance and Claims
The paper claims that uncertainty disentanglement is not only feasible in regression tasks but is potentially more stable and interpretable than in classification tasks. The results validate the theoretical distinction between the two uncertainty types:

Aleatoric uncertainty is confirmed as data-inherent, remaining constant regardless of model capacity or data volume.
Epistemic uncertainty is confirmed as model-dependent, decreasing as the model gains more exposure to data.

The authors conclude that Deep Ensembles offer the strongest raw predictive performance and robustness, while DropConnect provides the most responsive and interpretable epistemic trends, making it suitable for applications requiring precise confidence estimation. The study highlights the value of models that can quantify why they are uncertain (data noise vs. lack of knowledge), which is critical for high-stakes applications like age-restricted content enforcement.

Limitations
The authors note several constraints:

Small Subsets: Models trained on very small subsets (1% data) suffered from underfitting, rendering their uncertainty estimates unreliable.
Dataset Bias: The APPA-REAL dataset is skewed toward the 20–40 age range, potentially limiting generalization to children and the elderly.
Architectural Homogeneity: The study relied exclusively on DenseNet-121; uncertainty behaviors might differ with other architectures like ResNet or Vision Transformers.
Method Stability: Flipout demonstrated instability in its variational inference approximations, particularly at intermediate dataset sizes.

Technical Summary: Disentangling Model and Human Data Uncertainty in Apparent Facial Age Estimation

More like this