Face-recognition AI can score well overall while treating groups unevenly

Tests of three image models found that strong average performance did not remove disparities among race, gender and combined groups.

Editorial illustration — not from the study.

Researchers evaluated ResNet50, MobileNetV3 and the Data-efficient Image Transformer using pretrained weights and two facial-image datasets: FairFace, designed to provide balanced demographic groups, and the less controlled UTKFace dataset. They tested how well the models’ facial features supported image retrieval and demographic classification.

Uneven results across groups

A simple linear classifier consistently separated demographic information more effectively than one-nearest-neighbour retrieval on both datasets, showing that the pretrained facial representations contained demographic information that could be extracted with a basic classifier. ResNet50 had the highest overall classification performance, while MobileNetV3 produced more consistent subgroup results in several settings. Race–gender intersectional analysis found persistent demographic disparities even when overall accuracy improved. The study also evaluated fairness gaps, variation between groups and statistical significance, but the abstract does not report the numerical results of those tests.

Why average accuracy is not enough

Average accuracy can hide poorer performance for particular demographic groups. The findings support evaluating facial-recognition systems across individual and combined demographic groups rather than relying only on a single overall score.

What the tests show

The evidence comes from tests of three ImageNet-pretrained image models on the FairFace and UTKFace datasets, using retrieval and demographic-classification tasks. The study provides a framework for comparing representation quality and fairness, but results from these two datasets and these three models do not establish how all facial-recognition systems would perform in other datasets or real-world uses. The abstract also does not give numerical subgroup results or describe the practical consequences of the measured disparities.

A virtual reality test was linked to standard cognitive scores in older adults

A virtual reality cognitive screening tool showed moderate-to-strong associations with scores from an established assessment in older adults. The findings suggest that virtual reality could complement traditional screening, although the study included only 38 people who completed both tests.

4 hours ago·2 views

AI & Computing

AI scores for grant proposals were less reliable than expert reviews

Researchers compared scores from several large language models with expert ratings of 142 Swedish medical fellowship applications. The AI scores showed weaker agreement with expert judgments than the reviewers showed with one another, although the systems could still have a limited role in sorting applications.

21 hours ago·8 views

AI & Computing

High-mounted cameras spot pedestrians on the device at busy intersections

Researchers introduced a 13,314-image dataset for testing AI systems that view urban intersections from above. The results show that detection can run on edge devices such as the Jetson Orin Nano, although pedestrians remain harder to detect than vehicles and performance changes with scene conditions.

1 day ago·10 views