An old carny trick goes high tech — AI guesses ages better than we can

A facial recognition algorithm written by German researchers is pretty good at guessing both the age and ethnicity of faces in scanned pictures, though the scientists do not know why the software is, on average, more accurate than the human brain.

Researchers at Ruhr-Universität Bochum’s neural computation institute, say their graph-based slow-feature analysis algorithm has a mean absolute error of 3.41 years. They say that result is “a competitive performance for this challenging problem” and, in fact, is better than human results.

Assigning an ethnic group was easier still. The authors say the software’s probability of correctly guessing a subject’s origin is better than 99 percent despite the fact that the images’ average brightness was standardized.

Researchers Alberto N. Escalante-B. and Laurenz Wiskott report in the May 2020 journal Machine Learning that “slow feature analysis (SFA) is an unsupervised learning algorithm that extracts slowly varying features from a multi-dimensional time series.”

Versions of an SFA working with images by their nature needlessly discard otherwise useful information. To get around this and improve estimation accuracy, the authors have proposed an “extension called hierarchical information-preserving GSFA.”

Thousands of photos of faces of different ages were used in the experiment, according to a Ruhr-Universität Bochum release.

The scientists sorted the images by age, and the software picked through them for features that change slowly from face to face. Eye color and skin tone, for example, change abruptly from one face to another. On the other hand, wrinkles slowly grow.

Wiskott is quoted in the release saying, “We’re not quite sure what features our algorithm is looking for” because it has “learned to assess faces.”

Article Topics

age verification  |  AI  |  algorithms  |  artificial intelligence  |  biometrics  |  biometrics research  |  facial recognition  |  unsupervised learning