Reported by Cody Schmidt, class of 2025
This was the first event of the Computational Humanities Seminar series, which focuses on the role of technology in the social sciences. The series is organized by Jaehee Choi, Zhaojin Zheng, and Alice Xiang.
Professor Amir Fekrazad, a professor of economics from Texas A&M – San Antonio, presented his research on using artificial intelligence to estimate a person’s remaining lifespan on February 24th. Moderated by Professor Jaehee Choi, Professor Fekrazad detailed the process of creating such technology.
He began by posing his research question: “Can an AI model predict with adequate accuracy how many years a person has left to live solely using the person’s facial image?” He imagined a world where such technology is readily accessible, presenting a mock-up of a conversation with Siri where the user is given their estimated remaining life span through the iPhone’s facial scan technology.
Professor Fekrazad demonstrated the faultiness of human estimation, displaying an AI-generated image of a man. He asked an audience member to guess the fictional man’s age and remaining lifespan. The audience predicted the man at 40-50 years of age, with a full remaining lifespan of about 40 years, pointing to slight graying of the hair, faint wrinkling, and a healthy smile and complexion. Fekrazad claimed that an AI could perform better on this task, using their “black box” to bypass obvious factors and detect more subtle features, such as forehead size, eye shape, slight discoloration, or anything else that may be a discreet indicator of remaining life span.
To make these improvements in estimation a reality, Professor Fekrazad first identified the data and modeling needed for this technology. He gathered a novel dataset of 24,000 images from Wikidata and Wikipedia. They were of individuals who died of natural causes between 1990 and 2022, with the year the picture was taken, the birth year, and death year being recorded. The difference between the year of the picture and the death year was recorded as the remaining lifespan, or RL.
Prior to training the AI, the images underwent a series of changes. They were converted to black and white, given a horizontal alignment, and cropped the face to the center of the image. This ensured consistency when training the model. Regression analysis was then used to place a relationship between the image and the RL.
A deep convolutional neural network, an artificial neural network used in image and video recognition, was utilized. He fine-tuned an existing model to fit this task, training the AI with 70% of the collected data and using the remaining 30% for validation. Upon a random inspection of the data, only 2% were erroneous, which was deemed negligible.
The best model had a mean absolute error of 8.3 in the validation set, with higher accuracy when guessing the RL of older individuals. Professor Fekrazad was very satisfied with this model, pointing to the unpredictability and uncertainty of death in many cases. With these factors, and with this experiment being the first of its kind, these results are promising for the future of image and facial recognition, as well as machine learning.