Original title: Transformer-Based Multi-Task Learning with a Task-Aware Weighted Ensemble for Simultaneous Age Estimation and Gender Classification from Facial Images
Sümeyye Sarıateş, Erdal Özbay · September 2026
Guessing a person's age and gender from a photo is hard. Lighting, head angle, facial expressions, and aging change how faces look. Researchers wanted to solve both tasks at once using artificial intelligence vision transformers, which are computer models that process images. They built a system where a single core model inspects a face image to pull out useful visual patterns, and then two separate branches make the final guesses: one predicts age and the other predicts gender. They tested three different transformer models and combined their strengths using a specialized strategy called a Task-aware Weighted Ensemble. This method gives different weights to each model depending on the task. Overall, this combined approach achieved an average age error of 4.8905 years and a gender accuracy of 94.86%. It showed a statistically significant boost for guessing gender, but only a modest, non-significant improvement for guessing age.
Imagine three friends each trying to guess a person's age and gender from a blurry photo. One friend is good at spotting wrinkles, another at noticing jaw shape, and the third at seeing hairstyle. If you just average their guesses, you might get a muddled answer. Instead, you ask each friend to rate how confident they are for each guess, then trust the most confident one for age and the most confident one for gender separately. In this paper, the three vision transformers are like those friends, and the Task-aware Weighted Ensemble decides which transformer to trust most for each task.
Systems that estimate age or classify gender from faces are used in security, marketing, and healthcare. But doing both tasks separately is slow and ignores how they share useful visual clues. This work shows a single model can handle both at once, and a clever voting method between different vision transformers improves accuracy, especially for gender. That means faster, more reliable systems for real-world applications like age-restricted access or personalized services.
The abstract notes that the improvement in age estimation from TAWE over equal-weight averaging was modest and not statistically significant. It also only evaluates on the UTKFace dataset, which has 15,503 images, so results may not generalize to other datasets or real-world conditions. The work does not address fairness across different demographic groups or performance under extreme occlusions or low-quality images.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Transformer-Based Multi-Task Learning with a Task-Aware Weighted Ensemble for Simultaneous Age Estimation and Gender Classification from Facial Images — Sümeyye Sarıateş, Erdal Özbay — Sensors — doi:10.3390/s26185982
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.